@ngocsangairvds/vsaf 3.2.13 → 3.2.15

package/tools/vds-scripts/audit_orchestrator/src/vds_audit_orchestrator/engine/batch_processor.py

@@ -1,4776 +0,0 @@
-"""Batch Row Processor - Processes checklist rows in configurable batches.
-
-Provides:
-- Configurable batch sizes (FR-3.1)
-- Per-row and per-batch timeouts (FR-3.2)
-- Checkpoint/resume with full row results (FR-3.3)
-- Progress logging
-- Target filtering for dynamic analysis (FR-25)
-"""
-
-from __future__ import annotations
-
-import asyncio
-import contextlib
-import copy
-import inspect
-import json
-import math
-import os
-from collections.abc import Awaitable, Callable
-from dataclasses import dataclass, field
-from datetime import UTC, datetime
-from pathlib import Path
-
-# Import GroundingValidator and TargetSelection only for typing, avoid circular if possible
-from typing import TYPE_CHECKING, Any, cast
-
-from structlog import get_logger
-from vds_agent_core.llm.budget import BudgetExceededError
-from vds_agent_core.profiles import resolve_default_failover_profiles
-
-from vds_audit_orchestrator.engine.provider_failure_classifier import (
-    ProviderFailureClass,
-    ProviderFailureClassification,
-    ProviderHealthMemory,
-    RowFailoverContext,
-    TimeoutKind,
-)
-from vds_audit_orchestrator.engine.row_evaluator import RowEvaluator
-from vds_audit_orchestrator.engine.row_evaluator_types import RowEvaluationResult, requires_app_config_only
-from vds_audit_orchestrator.errors import AUDIT_ERROR_CODES
-from vds_audit_orchestrator.evidence.matcher import MatcherProtocol, RowEvidenceContext
-from vds_audit_orchestrator.models.checklist import (
-    VERIFICATION_REASON_EXCERPT_VERIFIED_IN_CONTEXT,
-    VERIFICATION_REASON_FALLBACK_REF_INHERITED,
-    EvidenceAnchor,
-    RowProvenance,
-    RowStatus,
-    ScoreBreakdown,
-    normalize_verification_reason,
-)
-from vds_audit_orchestrator.runtime_profiles import inherit_runtime_llm_policy
-
-if TYPE_CHECKING:
-    from vds_audit_orchestrator.engine.grounding_validator import GroundingValidator
-    from vds_audit_orchestrator.engine.target_selector import TargetSelection
-    from vds_audit_orchestrator.models.template import AuditCheck, AuditTemplate
-
-logger = get_logger()
-_TIMEOUT_PROGRESS_COUNTER_KEYS: tuple[str, ...] = (
-    "tool_calls",
-    "skill_calls",
-    "skill_execution_calls",
-    "skill_effective_calls",
-)
-
-
-def _convert_datetimes(obj: Any) -> Any:
-    """Recursively convert datetime objects to ISO strings for JSON serialization."""
-    if isinstance(obj, datetime):
-        return obj.isoformat()
-    if isinstance(obj, dict):
-        return {k: _convert_datetimes(v) for k, v in obj.items()}
-    if isinstance(obj, list):
-        return [_convert_datetimes(item) for item in obj]
-    return obj
-
-
-def _sanitize_retrieval_trace(trace: Any) -> dict[str, Any] | None:
-    """Return JSON-serializable retrieval trace dict or None."""
-    if not isinstance(trace, dict):
-        return None
-    normalized = _convert_datetimes(trace)
-    try:
-        json.dumps(normalized, ensure_ascii=False)
-    except TypeError:
-        return None
-    return normalized
-
-
-def _merge_retrieval_trace_payloads(
-    matcher_trace: dict[str, Any] | None,
-    row_trace: dict[str, Any] | None,
-) -> dict[str, Any] | None:
-    """Merge matcher + row retrieval traces without losing non-null matcher context."""
-    base = dict(matcher_trace or {})
-    overlay = dict(row_trace or {})
-    if not base and not overlay:
-        return None
-    if not base:
-        return overlay
-    if not overlay:
-        return base
-
-    merged = dict(base)
-    preserve_empty_overrides = bool(
-        (
-            overlay.get("app_config_only_enforced") if isinstance(overlay.get("app_config_only_enforced"), dict) else {}
-        ).get("applied")
-    ) or RowEvaluator._is_app_config_only_requirement(retrieval_trace=overlay)
-    for key, value in overlay.items():
-        if value is None:
-            continue
-        existing = merged.get(key)
-        if isinstance(value, list):
-            if not value and isinstance(existing, list) and existing and not preserve_empty_overrides:
-                continue
-            merged[key] = value
-            continue
-        if isinstance(value, dict):
-            if not value and isinstance(existing, dict) and existing and not preserve_empty_overrides:
-                continue
-            if not value and preserve_empty_overrides:
-                merged[key] = value
-                continue
-            if isinstance(existing, dict):
-                nested = dict(existing)
-                for nested_key, nested_value in value.items():
-                    if nested_value is None:
-                        continue
-                    prior_nested = nested.get(nested_key)
-                    if (
-                        isinstance(nested_value, list)
-                        and not nested_value
-                        and isinstance(prior_nested, list)
-                        and prior_nested
-                        and not preserve_empty_overrides
-                    ):
-                        continue
-                    if (
-                        isinstance(nested_value, dict)
-                        and not nested_value
-                        and isinstance(prior_nested, dict)
-                        and prior_nested
-                        and not preserve_empty_overrides
-                    ):
-                        continue
-                    if isinstance(nested_value, dict) and not nested_value and preserve_empty_overrides:
-                        nested[nested_key] = nested_value
-                        continue
-                    nested[nested_key] = nested_value
-                merged[key] = nested
-                continue
-            merged[key] = value
-            continue
-        merged[key] = value
-    return merged
-
-
-def _default_batch_error_retry_limit() -> int:
-    """Return batch-level error retry limit from env, defaulting to 1 (FR-145.2)."""
-    try:
-        return max(0, int(os.environ.get("VDS_AUDIT_BATCH_ERROR_RETRY_LIMIT", "2")))
-    except (ValueError, TypeError):
-        return 1
-
-
-# ---------------------------------------------------------------------------
-# Phase 146: Quota-aware recovery configuration (FR-146.6)
-# ---------------------------------------------------------------------------
-
-
-def _quota_wait_max_seconds() -> float:
-    """Return the maximum seconds to wait when all providers are exhausted (AC-146.6.3).
-
-    Env: VDS_AUDIT_QUOTA_WAIT_MAX_SECONDS, default 30.0.
-    Set to 0 to disable bounded wait-and-resume entirely.
-    """
-    try:
-        return max(0.0, float(os.environ.get("VDS_AUDIT_QUOTA_WAIT_MAX_SECONDS", "30")))
-    except (ValueError, TypeError):
-        return 30.0
-
-
-def _exhausted_provider_verification_limit() -> int:
-    """Return max verification retries for a provider already known to be exhausted (AC-146.5.3).
-
-    Env: VDS_AUDIT_EXHAUSTED_PROVIDER_RETRIES, default 0.
-    When 0, known-exhausted providers are skipped entirely during retry sweeps.
-    """
-    try:
-        return max(0, int(os.environ.get("VDS_AUDIT_EXHAUSTED_PROVIDER_RETRIES", "0")))
-    except (ValueError, TypeError):
-        return 0
-
-
-def _all_providers_exhausted_reason(
-    health_memory: ProviderHealthMemory,
-    *,
-    failover_profiles: list[str],
-) -> str | None:
-    """Return an exhaustion reason code if all eligible providers are unusable (AC-146.6.4).
-
-    Returns:
-        ``quota_all_providers_exhausted`` when at least one provider is in
-        cooldown/constrained/degraded state (temporary), ``None`` when at
-        least one provider is still dispatchable, or
-        ``provider_all_auth_blocked`` when every provider is auth-blocked
-        (permanent for the current run).
-    """
-    if not failover_profiles:
-        return None
-
-    any_dispatchable = False
-    any_temporary = False
-    for profile in failover_profiles:
-        state = health_memory.get_state(profile)
-        if state.is_auth_blocked():
-            continue
-        if state.is_in_cooldown() or state.should_skip_dispatch():
-            any_temporary = True
-            continue
-        any_dispatchable = True
-        break
-
-    if any_dispatchable:
-        return None
-    if any_temporary:
-        return "quota_all_providers_exhausted"
-    # All auth-blocked — permanent for this run
-    if failover_profiles:
-        return "provider_all_auth_blocked"
-    return None
-
-
-@dataclass
-class BatchConfig:
-    """Configuration for batch processing (FR-3.1, FR-3.2, FR-25, FR-188, FR-145.2).
-
-    Attributes:
-        batch_size: Number of rows per batch (default 5).
-        row_timeout_ms: Timeout per row in milliseconds (default 45s).
-        row_progress_lease_seconds: Renewable liveness lease window in seconds.
-            Default 0 keeps legacy timeout-extension semantics.
-        row_stall_detection_seconds: No-progress stall threshold in seconds when
-            lease mode is enabled.
-        row_absolute_timeout_ms: Hard safety cap per row when lease mode is enabled.
-        batch_timeout_ms: Optional batch timeout (default: batch_size * effective row timeout ceiling + overhead).
-        batch_overhead_ms: Slack time for formatting + bookkeeping (default 15s).
-        batch_error_retry_limit: Max post-pass retry attempts for ERROR rows (FR-145.2).
-            Env: VDS_AUDIT_BATCH_ERROR_RETRY_LIMIT, default 2. Set 0 to disable.
-        checkpoint_dir: Directory for checkpoint files.
-        resume_from_checkpoint: Whether to resume from existing checkpoint.
-        target_selection: Optional target selection for dynamic analysis (FR-25).
-        force_refresh_targets: Re-evaluate targeted rows even if checkpointed (FR-25).
-        row_concurrency: Max rows dispatched in parallel within a batch (FR-188).
-            Default 1 (sequential, safe). Values >1 enable asyncio.gather dispatch
-            with a semaphore-bounded concurrency limit.
-    """
-
-    batch_size: int = 5  # FR-3.1: Default 5-10 rows
-    row_timeout_ms: int = 360_000  # Generous: 6 minutes per row (was 120s; profile YAML may override to 600s+)
-    row_progress_lease_seconds: int = 0  # FR-112.1: 0 preserves legacy timeout-extension mode
-    row_stall_detection_seconds: int = 60  # FR-112.1: stall threshold (was 30s)
-    row_absolute_timeout_ms: int = 1_200_000  # Generous: 20-minute hard safety cap (was 600s)
-    row_timeout_progress_extension_enabled: bool = True  # FR-72/73 follow-up: adaptive extension on effective progress
-    row_timeout_progress_retry_attempts: int = 5  # Generous: 5 extension attempts (was 2)
-    row_timeout_progress_extension_ms: int = 60_000  # Generous: 60s per extension (was 30s)
-    row_timeout_progress_max_ms: int = 600_000  # Generous: 10-minute ceiling after extensions (was 240s)
-    batch_timeout_ms: int | None = None  # Default: batch_size * effective row timeout ceiling + overhead
-    batch_overhead_ms: int = 15_000  # Slack for formatting + bookkeeping
-    batch_error_retry_limit: int = field(default_factory=_default_batch_error_retry_limit)  # FR-145.2
-    checkpoint_dir: Path | None = None
-    resume_from_checkpoint: bool = False
-    # FR-25: Dynamic Analysis Support - Target filtering
-    target_selection: TargetSelection | None = None
-    run_context: dict[str, Any] = field(default_factory=dict)
-    force_refresh_targets: bool = True  # Re-evaluate targeted rows even if checkpointed
-    # FR-185: Repo-type scoped N/A indices — skip LLM calls entirely for these rows
-    scoped_na_indices: set[int] = field(default_factory=set)
-    # FR-188: Parallel row dispatch within a batch
-    row_concurrency: int = 4  # Phase 158 default; >1 enables asyncio.gather dispatch
-    # FR-128.3 / Phase 158: adaptive concurrency is enabled by default.
-    adaptive_row_concurrency: bool = True
-    adaptive_row_concurrency_min: int = 2
-    adaptive_row_concurrency_max: int = 8
-    adaptive_row_concurrency_increase_threshold_ms: int = 1000
-    adaptive_row_concurrency_decrease_threshold_ms: int = 5000
-
-
-# ---------------------------------------------------------------------------
-# FR-145.2: Resilience budget enforcement
-# ---------------------------------------------------------------------------
-
-# Reason codes in retry_metadata that indicate a non-provider failure.
-# These rows will not benefit from retrying with a different provider.
-_NON_FAILOVERABLE_REASONS: frozenset[str] = frozenset(
-    {
-        ProviderFailureClass.TERMINAL_AUTH.value,
-        ProviderFailureClass.NON_PROVIDER_BUG.value,
-        "terminal_auth",
-        "non_provider_bug",
-        "schema_validation",
-        "template_mismatch",
-        "row_backend_invalid_env_value",
-        "target_resolution",
-    }
-)
-
-
-@dataclass
-class ResilienceBudget:
-    """Bounded retry budget shared across post-pass sweeps (FR-145.2, NFR-145.2).
-
-    The budget is generous and error-driven with a generous floor:
-        total = max(error_count * 3, max(15, ceil(total_rows * 0.35)))
-
-    For a 55-row batch with 12 ERRORs: max(36, max(15, 20)) = 36 retries.
-
-    Args:
-        total_rows: Total rows in the primary pass.
-        error_count: Number of ERROR rows from the primary pass.
-    """
-
-    total_rows: int
-    error_count: int
-    calls_used: int = 0
-
-    @property
-    def total(self) -> int:
-        """Return total resilience budget calls allowed."""
-        return max(self.error_count * 3, 15, math.ceil(self.total_rows * 0.35))
-
-    @property
-    def remaining(self) -> int:
-        return max(0, self.total - self.calls_used)
-
-    def consume(self) -> bool:
-        """Consume one budget slot. Returns False if budget exhausted."""
-        if self.calls_used >= self.total:
-            return False
-        self.calls_used += 1
-        return True
-
-
-@dataclass
-class _ResilienceRetryStats:
-    """Mutable accumulator for retry sweep statistics (FR-145.7).
-
-    Passed by reference into ``_retry_error_rows`` so the caller can read
-    the final counts without the method returning a tuple.
-    """
-
-    error_rows_before_retry: int = 0
-    retry_attempts: int = 0
-    retry_successes: int = 0
-    budget_used: int = 0
-    budget_total: int = 0
-    failover_hops: int = 0
-    providers_used: list[str] = field(default_factory=list)
-    quota_exhausted_rows: int = 0
-    quota_wait_attempted_rows: int = 0
-    quota_wait_recovered_rows: int = 0
-    # Phase 166: Prompt-level retry/failover telemetry
-    prompt_retry_attempts: int = 0
-    prompt_retry_recoveries: int = 0
-    prompt_failover_attempts: int = 0
-    prompt_failover_recoveries: int = 0
-    prompt_failover_exhausted: int = 0
-    synthesis_fallback_count: int = 0
-
-
-@dataclass
-class BatchResult:
-    """Result of processing a single batch."""
-
-    batch_index: int
-    start_row: int
-    end_row: int
-    results: list[RowEvaluationResult]
-    duration_ms: int
-    success_count: int
-    error_count: int
-    timed_out: bool = False
-    skipped_count: int = 0  # FR-25: Rows skipped due to targeting
-    postproc_overlap_ms: int = 0  # FR-193: Post-processing streaming overlap
-
-
-class BatchBudgetExceededError(RuntimeError):
-    """Raised when strict budget mode aborts a batch mid-execution."""
-
-    def __init__(
-        self,
-        *,
-        message: str,
-        context: dict[str, Any],
-        partial_results: list[RowEvaluationResult],
-        success_count: int,
-        error_count: int,
-        skipped_count: int,
-    ) -> None:
-        super().__init__(message)
-        self.context = context
-        self.partial_results = partial_results
-        self.success_count = success_count
-        self.error_count = error_count
-        self.skipped_count = skipped_count
-
-
-@dataclass(frozen=True)
-class BatchProgress:
-    """Progress snapshot emitted after a processed batch."""
-
-    batch_index: int
-    batch_count: int
-    start_row: int
-    end_row: int
-    total_rows: int
-    completed_rows: int
-    processed_batches: int
-    skipped_batches: int
-    success_count: int
-    error_count: int
-    skipped_count: int
-    duration_ms: int
-    timed_out: bool
-    emission_kind: str = "batch"
-    current_row_id: str | None = None
-    current_check_id: str | None = None
-
-
-BatchProgressCallback = Callable[[BatchProgress, list[RowEvaluationResult]], Awaitable[None] | None]
-RowProgressCallback = Callable[
-    [int, int, int, list[RowEvaluationResult], int, int, int, RowEvaluationResult | None, str],
-    Awaitable[None] | None,
-]
-
-
-@dataclass
-class BatchCheckpoint:
-    """Checkpoint for resumable processing (FR-3.3).
-
-    CRITICAL: Stores full row results to enable complete reconstruction on resume.
-    """
-
-    thread_id: str
-    total_rows: int
-    template_hash: str  # FR-3.3: For cache invalidation
-    evidence_hash: str  # FR-3.3: For cache invalidation
-    completed_batches: list[int]
-    # FR-3.3: Row ID -> serialized RowEvaluationResult (full results, not just status)
-    results_by_row_id: dict[str, dict[str, Any]]
-    last_updated: datetime = field(default_factory=lambda: datetime.now(UTC))
-
-    def save(self, path: Path) -> None:
-        """Save checkpoint to file."""
-        data = {
-            "thread_id": self.thread_id,
-            "total_rows": self.total_rows,
-            "template_hash": self.template_hash,
-            "evidence_hash": self.evidence_hash,
-            "completed_batches": self.completed_batches,
-            "results_by_row_id": self.results_by_row_id,
-            "last_updated": self.last_updated.isoformat(),
-        }
-        path.parent.mkdir(parents=True, exist_ok=True)
-        path.write_text(json.dumps(data, indent=2, ensure_ascii=False))
-
-    @classmethod
-    def load(cls, path: Path) -> BatchCheckpoint | None:
-        """Load checkpoint from file."""
-        if not path.exists():
-            return None
-        try:
-            data = json.loads(path.read_text())
-            return cls(
-                thread_id=data["thread_id"],
-                total_rows=data["total_rows"],
-                template_hash=data.get("template_hash", ""),
-                evidence_hash=data.get("evidence_hash", ""),
-                completed_batches=data["completed_batches"],
-                results_by_row_id=data.get("results_by_row_id", {}),
-                last_updated=datetime.fromisoformat(data["last_updated"]),
-            )
-        except Exception as e:
-            logger.warning("checkpoint_load_failed", error=str(e))
-            return None
-
-    @staticmethod
-    def serialize_row_result(result: RowEvaluationResult) -> dict[str, Any]:
-        """Serialize a RowEvaluationResult for checkpoint storage (FR-3.3)."""
-        # Basic serialization - assumes RowEvaluationResult and children have model_dump or similar
-        # Since RowEvaluationResult is Pydantic model (usually), model_dump is available
-        # But let's be safe and use json-compatible dict
-
-        # Helper to safely dump models
-        def safe_dump(obj: Any) -> Any:
-            if obj is None:
-                return None
-            if hasattr(obj, "model_dump"):
-                dumped = obj.model_dump()
-                # Convert datetime objects to ISO strings
-                return _convert_datetimes(dumped)
-            if hasattr(obj, "dict"):
-                dumped = obj.dict()
-                return _convert_datetimes(dumped)
-            if isinstance(obj, datetime):
-                return obj.isoformat()
-            return obj
-
-        retrieval_trace = _sanitize_retrieval_trace(result.retrieval_trace)
-
-        return {
-            "row_id": result.row_id,
-            "check_id": result.check_id,
-            "status": result.status.value if hasattr(result.status, "value") else str(result.status),
-            "score": result.score,
-            "score_breakdown": safe_dump(result.score_breakdown),
-            "reason": result.reason,
-            "finding": result.finding,
-            "evidence_anchors": [safe_dump(e) for e in (result.evidence_anchors or [])],
-            "provenance": safe_dump(result.provenance),
-            "severity": result.severity,
-            "priority": result.priority,
-            "effort": result.effort,
-            "recommendation": result.recommendation,
-            "fix_suggestions": result.fix_suggestions,
-            "score_1_5": result.score_1_5,
-            "cache_hit": result.cache_hit,
-            "error_message": result.error_message,
-            "retry_count": result.retry_count,
-            "retry_metadata": result.retry_metadata,
-            "retrieval_trace": retrieval_trace,
-            "coverage_requirements": result.coverage_requirements,
-            "coverage_code": result.coverage_code,
-            "coverage_docs": result.coverage_docs,
-            "coverage_confidence": result.coverage_confidence,
-            "coverage_deterministic_downgrade_flag": result.coverage_deterministic_downgrade_flag,
-        }
-
-    @staticmethod
-    def deserialize_row_result(data: dict[str, Any]) -> RowEvaluationResult:
-        """Deserialize a RowEvaluationResult from checkpoint (FR-3.3)."""
-        # We need to reconstruct the objects. This requires importing the types.
-        # This is best effort reconstruction for resumption.
-        from vds_audit_orchestrator.engine.row_evaluator_types import RowEvaluationResult
-        from vds_audit_orchestrator.models.checklist import (
-            EvidenceAnchor,
-            RowProvenance,
-            RowStatus,
-            ScoreBreakdown,
-        )
-
-        # Handle status enum
-        status_val = data.get("status", "ERROR")
-        try:
-            status = RowStatus(status_val)
-        except ValueError:
-            status = RowStatus.ERROR
-
-        # Reconstruct complex objects
-        score_breakdown = (
-            ScoreBreakdown(**data.get("score_breakdown", {}))
-            if data.get("score_breakdown")
-            else ScoreBreakdown.compute(data.get("score", 0.0))
-        )
-
-        anchors = []
-        for a_data in data.get("evidence_anchors", []):
-            anchors.append(EvidenceAnchor(**a_data))
-
-        provenance = RowProvenance(**data.get("provenance", {})) if data.get("provenance") else None
-
-        # Remove keys that are not fields of RowEvaluationResult or handle extra fields
-        # Ideally verify against RowEvaluationResult.__init__ args
-
-        return RowEvaluationResult(
-            row_id=data["row_id"],
-            check_id=data["check_id"],
-            status=status,
-            score=data.get("score", 0.0),
-            score_breakdown=score_breakdown,
-            reason=data.get("reason") or "",
-            reasoning=data.get("reasoning") or "",
-            finding=data.get("finding") or "",
-            evidence_anchors=anchors,
-            provenance=provenance,  # type: ignore[arg-type]
-            severity=data.get("severity") or "",
-            priority=data.get("priority") or "",
-            effort=data.get("effort") or "",
-            recommendation=data.get("recommendation") or "",
-            fix_suggestions=data.get("fix_suggestions") or [],
-            score_1_5=data.get("score_1_5"),
-            cache_hit=data.get("cache_hit", False),
-            error_message=data.get("error_message"),
-            retry_count=max(0, int(data.get("retry_count", 0))),
-            retry_metadata=dict(data.get("retry_metadata") or {}),
-            retrieval_trace=_sanitize_retrieval_trace(data.get("retrieval_trace")),
-            coverage_requirements=data.get("coverage_requirements"),
-            coverage_code=data.get("coverage_code"),
-            coverage_docs=data.get("coverage_docs"),
-            coverage_confidence=data.get("coverage_confidence"),
-            coverage_deterministic_downgrade_flag=bool(data.get("coverage_deterministic_downgrade_flag", False)),
-        )
-
-
-class BatchRowProcessor:
-    """Processes checklist rows in batches with checkpoint support.
-
-    Supports targeted row execution (FR-25) where only specific rows are
-    evaluated while preserving checkpoint state for non-targeted rows.
-    """
-
-    def __init__(
-        self,
-        template: AuditTemplate,
-        evaluator: RowEvaluator,
-        evidence_matcher: MatcherProtocol,
-        config: BatchConfig,
-        *,
-        grounding_validator: GroundingValidator | None = None,
-    ):
-        """Initialize batch processor.
-
-        Args:
-            template: Audit template with checks.
-            evaluator: Row evaluator instance.
-            evidence_matcher: Evidence matcher for per-row context.
-            config: Batch processing configuration.
-            grounding_validator: Optional validator for FR-7.
-        """
-        self.template = template
-        self.evaluator = evaluator
-        self.evidence_matcher = evidence_matcher
-        self.config = config
-        self.grounding_validator = grounding_validator
-
-        # Flatten checks for easier batch processing
-        self._checks: list[tuple[str, AuditCheck]] = []
-        for section in template.sections:
-            for check in section.checks:
-                self._checks.append((section.id, check))
-        self.last_progress: BatchProgress | None = None
-        self.last_execution_summary: dict[str, int] = {
-            "processed_batches": 0,
-            "skipped_batches": 0,
-            "total_results": 0,
-            "batch_count": 0,
-        }
-        # FR-145.7: Per-run resilience telemetry; reset at process_all start.
-        self.last_resilience_summary: dict[str, Any] | None = None
-        self._resilience_stats_accumulator: list[_ResilienceRetryStats] = []
-        self._batch_match_cache: dict[tuple[str, str | None, str, bool], RowEvidenceContext] = {}
-        self._batch_match_inflight: dict[tuple[str, str | None, str, bool], asyncio.Future[RowEvidenceContext]] = {}
-        self._batch_match_lock = asyncio.Lock()
-        self._adaptive_row_concurrency_current = max(1, int(self.config.row_concurrency or 1))
-
-    @staticmethod
-    def _preflight_provider_checks_from_run_context(run_context: dict[str, Any] | None) -> dict[str, dict[str, Any]]:
-        if not isinstance(run_context, dict):
-            return {}
-        raw_checks = run_context.get("profile_availability_checks")
-        if not isinstance(raw_checks, list):
-            return {}
-
-        normalized: dict[str, dict[str, Any]] = {}
-        for item in raw_checks:
-            if not isinstance(item, dict):
-                continue
-            profile_name = str(item.get("profile") or "").strip()
-            if not profile_name:
-                continue
-            normalized[profile_name] = dict(item)
-        return normalized
-
-    async def _seed_batch_health_memory_from_preflight(self) -> None:
-        runtime_context = self.config.run_context if isinstance(self.config.run_context, dict) else {}
-        preflight_checks = self._preflight_provider_checks_from_run_context(runtime_context)
-        if not preflight_checks:
-            return
-
-        for provider_name, probe in preflight_checks.items():
-            await self._batch_health_memory.apply_preflight_probe(provider_name, probe)
-
-    @staticmethod
-    async def _record_runtime_provider_result(
-        health_memory: ProviderHealthMemory | None,
-        *,
-        provider_name: str | None,
-        result: RowEvaluationResult,
-    ) -> None:
-        if health_memory is None:
-            return
-
-        normalized_provider = str(provider_name or "").strip()
-        if not normalized_provider:
-            return
-
-        retry_metadata = dict(result.retry_metadata or {}) if isinstance(result.retry_metadata, dict) else {}
-        if result.status == RowStatus.ERROR:
-            reason_code = str(retry_metadata.get("reason_code") or "").strip().lower()
-            if not reason_code:
-                return
-            failure_class = {
-                "provider_authentication_failed": ProviderFailureClass.TERMINAL_AUTH,
-                "provider_authorization_failed": ProviderFailureClass.TERMINAL_AUTH,
-                "provider_rate_limited": ProviderFailureClass.QUOTA_OR_CAPACITY,
-                "provider_server_error": ProviderFailureClass.RETRYABLE_TRANSIENT,
-                "provider_transient_error": ProviderFailureClass.RETRYABLE_TRANSIENT,
-                "provider_http_error": ProviderFailureClass.RETRYABLE_TRANSIENT,
-            }.get(reason_code)
-            if failure_class is None:
-                return
-            retry_after_raw = retry_metadata.get("retry_after_seconds")
-            retry_after_seconds: float | None
-            try:
-                retry_after_seconds = float(retry_after_raw) if retry_after_raw is not None else None
-            except (TypeError, ValueError):
-                retry_after_seconds = None
-
-            await health_memory.record_failure(
-                normalized_provider,
-                classification=ProviderFailureClassification(
-                    failure_class=failure_class,
-                    is_failoverable=failure_class != ProviderFailureClass.TERMINAL_AUTH,
-                    retry_after_seconds=retry_after_seconds,
-                    raw_status_code=None,
-                    raw_message=reason_code,
-                    classification_reason=reason_code,
-                ),
-            )
-            return
-
-        await health_memory.record_success(normalized_provider)
-
-    # ------------------------------------------------------------------
-    # Phase 146: Bounded all-providers-exhausted recovery (FR-146.6)
-    # ------------------------------------------------------------------
-
-    @staticmethod
-    async def _bounded_wait_for_provider_recovery(
-        health_memory: ProviderHealthMemory,
-        *,
-        failover_profiles: list[str],
-        max_wait_seconds: float,
-        poll_interval_seconds: float = 2.0,
-        row_id: str | None = None,
-        check_id: str | None = None,
-    ) -> str | None:
-        """Wait bounded time for a provider to recover from cooldown/exhaustion (AC-146.6.1).
-
-        Uses ``nearest_recovery_seconds()`` from the shared provider health state
-        to estimate the earliest possible recovery.  The wait is capped by
-        ``max_wait_seconds`` (from ``VDS_AUDIT_QUOTA_WAIT_MAX_SECONDS``).
-
-        Returns:
-            A recovered profile name, or ``None`` if no provider became eligible
-            within the cap.
-
-        Telemetry:
-            Emits structured log events for wait-start, provider-recovered, and
-            wait-expired boundaries so operators can diagnose quota timing.
-        """
-        import time as _time
-
-        if max_wait_seconds <= 0 or not failover_profiles:
-            return None
-
-        # Check if any provider is only temporarily blocked (not auth-blocked).
-        has_temporary_blockage = any(
-            not health_memory.get_state(p).is_auth_blocked() and health_memory.get_state(p).should_skip_dispatch()
-            for p in failover_profiles
-        )
-        if not has_temporary_blockage:
-            return None
-
-        # Determine how long to wait: min of configured cap and nearest recovery hint.
-        recovery_hint = health_memory.nearest_recovery_seconds()
-        if recovery_hint is not None and recovery_hint > 0:
-            wait_target = min(max_wait_seconds, recovery_hint + 1.0)
-        else:
-            wait_target = max_wait_seconds
-
-        logger.info(
-            "quota_bounded_wait_started",
-            row_id=row_id,
-            check_id=check_id,
-            max_wait_seconds=max_wait_seconds,
-            recovery_hint_seconds=recovery_hint,
-            wait_target_seconds=wait_target,
-            failover_profiles=failover_profiles,
-        )
-
-        deadline = _time.monotonic() + wait_target
-        recovered_profile: str | None = None
-
-        while _time.monotonic() < deadline:
-            remaining = deadline - _time.monotonic()
-            if remaining <= 0:
-                break
-            sleep_seconds = min(poll_interval_seconds, remaining)
-            await asyncio.sleep(sleep_seconds)
-
-            # Check if any provider has become eligible.
-            for candidate in failover_profiles:
-                state = health_memory.get_state(candidate)
-                if not state.is_auth_blocked() and not state.should_skip_dispatch():
-                    recovered_profile = candidate
-                    break
-            if recovered_profile is not None:
-                break
-
-        if recovered_profile is not None:
-            logger.info(
-                "quota_bounded_wait_provider_recovered",
-                row_id=row_id,
-                check_id=check_id,
-                recovered_provider=recovered_profile,
-                elapsed_seconds=wait_target - (deadline - _time.monotonic()),
-            )
-        else:
-            logger.warning(
-                "quota_bounded_wait_expired",
-                row_id=row_id,
-                check_id=check_id,
-                max_wait_seconds=max_wait_seconds,
-                failover_profiles=failover_profiles,
-            )
-
-        return recovered_profile
-
-    @staticmethod
-    def _annotate_quota_exhaustion(
-        result: RowEvaluationResult,
-        *,
-        health_memory: ProviderHealthMemory,
-        failover_profiles: list[str],
-        wait_attempted: bool = False,
-        wait_expired: bool = False,
-    ) -> RowEvaluationResult:
-        """Annotate a row result with quota-exhaustion metadata (AC-146.6.4).
-
-        Mutates ``retry_metadata`` in-place and returns the same result.
-        """
-        reason = _all_providers_exhausted_reason(health_memory, failover_profiles=failover_profiles)
-        if reason is None:
-            return result
-
-        metadata = dict(result.retry_metadata or {})
-        metadata["quota_all_providers_exhausted"] = True
-        metadata["quota_exhaustion_reason"] = reason
-        metadata["quota_exhaustion_wait_attempted"] = wait_attempted
-        metadata["quota_exhaustion_wait_expired"] = wait_expired
-
-        # Attach per-provider skip reasons for telemetry clarity (AC-146.7.2).
-        provider_skip_reasons: dict[str, str] = {}
-        for profile in failover_profiles:
-            state = health_memory.get_state(profile)
-            if state.is_auth_blocked():
-                provider_skip_reasons[profile] = "auth_blocked"
-            elif state.is_in_cooldown():
-                provider_skip_reasons[profile] = "cooldown"
-            elif state.rate_limit.request_capacity_known_empty():
-                provider_skip_reasons[profile] = "request_quota_empty"
-            elif state.rate_limit.token_capacity_known_empty():
-                provider_skip_reasons[profile] = "token_quota_empty"
-            elif state.should_skip_dispatch():
-                provider_skip_reasons[profile] = "degraded"
-            else:
-                provider_skip_reasons[profile] = "eligible"
-        metadata["quota_provider_skip_reasons"] = provider_skip_reasons
-        result.retry_metadata = metadata
-        return result
-
-    @staticmethod
-    def _build_match_requirement_text(check: AuditCheck) -> str:
-        """Augment matcher query with canonical checklist hints."""
-        base = (check.description or check.name or "").strip()
-        config = check.check_config if isinstance(check.check_config, dict) else {}
-        extras = [
-            str(config.get("confluence_rule") or "").strip(),
-            str(config.get("confluence_main") or "").strip(),
-            str(config.get("confluence_eval_guide") or "").strip(),
-            str(config.get("confluence_evidence_hint") or "").strip(),
-            str(config.get("evidence_required") or "").strip(),
-        ]
-        suffix = " ".join(part for part in extras if part)
-        if base and suffix:
-            return f"{base} {suffix}"
-        return base or suffix
-
-    @staticmethod
-    def _row_requires_app_config_only(check: AuditCheck) -> bool:
-        """Return True when row guidance explicitly restricts retrieval to app-config."""
-        config = check.check_config if isinstance(check.check_config, dict) else {}
-        return requires_app_config_only(
-            check.description or check.name or "",
-            config.get("confluence_rule"),
-            config.get("confluence_main"),
-            config.get("confluence_eval_guide"),
-            config.get("confluence_evidence_hint"),
-            config.get("detailed_guidance"),
-            config.get("confluence_notes"),
-            config.get("evidence_required"),
-        )
-
-    @staticmethod
-    def _empty_evidence_context(*, row_id: str, requirement_text: str) -> RowEvidenceContext:
-        """Return an empty evidence context for rows that must bypass matcher retrieval."""
-        return RowEvidenceContext(
-            row_id=row_id,
-            requirement_text=requirement_text,
-            matched_docs=[],
-            matched_code=[],
-            evidence_refs=[],
-            total_chars=0,
-            truncated=False,
-            retrieval_trace=None,
-        )
-
-    @staticmethod
-    def _clone_evidence_context(
-        *, context: RowEvidenceContext, row_id: str, requirement_text: str
-    ) -> RowEvidenceContext:
-        """Clone matcher context so cached retrieval can be reused safely per row."""
-        return RowEvidenceContext(
-            row_id=row_id,
-            requirement_text=requirement_text,
-            matched_docs=copy.deepcopy(getattr(context, "matched_docs", []) or []),
-            matched_code=copy.deepcopy(getattr(context, "matched_code", []) or []),
-            evidence_refs=list(getattr(context, "evidence_refs", []) or []),
-            total_chars=int(getattr(context, "total_chars", 0) or 0),
-            truncated=bool(getattr(context, "truncated", False)),
-            retrieval_trace=copy.deepcopy(getattr(context, "retrieval_trace", None)),
-        )
-
-    async def _get_or_match_evidence_context(
-        self,
-        *,
-        row_id: str,
-        requirement_text: str,
-        section_id: str | None,
-        app_config_only: bool,
-    ) -> RowEvidenceContext:
-        """Return row evidence context with in-batch dedup for parallel rows."""
-        if app_config_only:
-            return self._empty_evidence_context(row_id=row_id, requirement_text=requirement_text)
-
-        cache_key = (
-            requirement_text,
-            str(section_id) if section_id is not None else None,
-            str(getattr(self.evidence_matcher, "retrieval_mode", "unknown") or "unknown"),
-            False,
-        )
-
-        async with self._batch_match_lock:
-            cached = self._batch_match_cache.get(cache_key)
-            if cached is not None:
-                return self._clone_evidence_context(
-                    context=cached,
-                    row_id=row_id,
-                    requirement_text=requirement_text,
-                )
-            inflight = self._batch_match_inflight.get(cache_key)
-            if inflight is None:
-                loop = asyncio.get_running_loop()
-                inflight = loop.create_future()
-                self._batch_match_inflight[cache_key] = inflight
-                owner = True
-            else:
-                owner = False
-
-        if owner:
-            try:
-                matched = await asyncio.to_thread(
-                    self.evidence_matcher.match_row,
-                    row_id=row_id,
-                    requirement_text=requirement_text,
-                    section_id=section_id,
-                )
-                canonical = self._clone_evidence_context(
-                    context=matched,
-                    row_id="__shared__",
-                    requirement_text=requirement_text,
-                )
-                async with self._batch_match_lock:
-                    self._batch_match_cache[cache_key] = canonical
-                    pending = self._batch_match_inflight.pop(cache_key, None)
-                    if pending is not None and not pending.done():
-                        pending.set_result(canonical)
-                return self._clone_evidence_context(
-                    context=canonical,
-                    row_id=row_id,
-                    requirement_text=requirement_text,
-                )
-            except Exception as exc:
-                async with self._batch_match_lock:
-                    pending = self._batch_match_inflight.pop(cache_key, None)
-                    if pending is not None and not pending.done():
-                        pending.set_exception(exc)
-                raise
-
-        shared_context = await inflight
-        return self._clone_evidence_context(
-            context=shared_context,
-            row_id=row_id,
-            requirement_text=requirement_text,
-        )
-
-    def _effective_row_concurrency(self) -> int:
-        """Return current row concurrency after optional adaptive adjustment."""
-        if not self.config.adaptive_row_concurrency:
-            return max(1, int(self.config.row_concurrency or 1))
-        configured = max(1, int(self.config.row_concurrency or 1))
-        lower = max(1, int(self.config.adaptive_row_concurrency_min or 1))
-        upper = max(lower, int(self.config.adaptive_row_concurrency_max or configured))
-        return max(lower, min(upper, int(self._adaptive_row_concurrency_current or configured)))
-
-    def _repo_row_distribution_profiles(self) -> list[str]:
-        runtime_context = self.config.run_context if isinstance(self.config.run_context, dict) else {}
-        raw_profiles = runtime_context.get("repo_distribution_profiles")
-        if not isinstance(raw_profiles, (list, tuple)):
-            return []
-        ordered: list[str] = []
-        seen: set[str] = set()
-        for item in raw_profiles:
-            normalized = str(item or "").strip()
-            if not normalized or normalized in seen:
-                continue
-            seen.add(normalized)
-            ordered.append(normalized)
-        return ordered
-
-    def _repo_row_distribution_enabled(self) -> bool:
-        runtime_context = self.config.run_context if isinstance(self.config.run_context, dict) else {}
-        mode = str(runtime_context.get("repo_profile_execution_mode") or "single").strip().lower()
-        return mode == "distributed" and len(self._repo_row_distribution_profiles()) > 1
-
-    def _assigned_repo_profile_for_row(self, row_index: int) -> str | None:
-        profiles = self._repo_row_distribution_profiles()
-        if not self._repo_row_distribution_enabled() or not profiles:
-            return None
-        return profiles[row_index % len(profiles)]
-
-    def _update_adaptive_row_concurrency(self, batch_result: BatchResult) -> None:
-        """Adjust opt-in row concurrency between batches based on observed batch health."""
-        if not self.config.adaptive_row_concurrency:
-            return
-
-        current = self._effective_row_concurrency()
-        lower = max(1, int(self.config.adaptive_row_concurrency_min or 1))
-        upper = max(lower, int(self.config.adaptive_row_concurrency_max or current))
-        increase_threshold_ms = max(1, int(self.config.adaptive_row_concurrency_increase_threshold_ms or 1000))
-        decrease_threshold_ms = max(
-            increase_threshold_ms, int(self.config.adaptive_row_concurrency_decrease_threshold_ms or 5000)
-        )
-
-        next_value = current
-        reason = None
-        if batch_result.timed_out or batch_result.error_count > 0 or batch_result.duration_ms >= decrease_threshold_ms:
-            if current > lower:
-                next_value = current - 1
-                reason = "batch_pressure"
-        elif (
-            batch_result.duration_ms <= increase_threshold_ms
-            and batch_result.error_count == 0
-            and not batch_result.timed_out
-        ) and current < upper:
-            next_value = current + 1
-            reason = "batch_healthy"
-
-        self._adaptive_row_concurrency_current = next_value
-        if reason is not None and next_value != current:
-            logger.info(
-                "adaptive_row_concurrency_adjusted",
-                previous=current,
-                current=next_value,
-                reason=reason,
-                batch_index=batch_result.batch_index,
-                batch_duration_ms=batch_result.duration_ms,
-                batch_error_count=batch_result.error_count,
-                batch_timed_out=batch_result.timed_out,
-                **self.config.run_context,
-            )
-
-    def _is_row_targeted(self, row_index: int, check_id: str) -> bool:
-        """Check if a row is targeted for evaluation (FR-25).
-
-        Args:
-            row_index: 0-based row index.
-            check_id: The check ID for this row.
-
-        Returns:
-            True if the row should be evaluated, False if it should be skipped.
-            Returns True if no target selection is configured (process all rows).
-        """
-        if self.config.target_selection is None:
-            return True
-
-        # Check if row index is in target set
-        if self.config.target_selection.matches_row_index(row_index):
-            return True
-
-        # Check if check ID is in target set (for CHECK_IDS type)
-        return bool(self.config.target_selection.matches_check_id(check_id.upper()))
-
-    def _batch_has_targeted_rows(self, start_idx: int, end_idx: int) -> bool:
-        """Check if a batch contains any targeted rows (FR-25.2).
-
-        Args:
-            start_idx: Start index of the batch (inclusive).
-            end_idx: End index of the batch (exclusive).
-
-        Returns:
-            True if the batch has at least one targeted row.
-            Returns True if no target selection is configured.
-        """
-        if self.config.target_selection is None:
-            return True
-
-        for idx in range(start_idx, end_idx):
-            _, check = self._checks[idx]
-            if self._is_row_targeted(idx, check.id):
-                return True
-        return False
-
-    def _should_emit_row_progress_callback(
-        self,
-        current_result: RowEvaluationResult | None,
-        outcome: str,
-    ) -> bool:
-        """Suppress noisy row-level progress events for skipped non-target rows."""
-        return not (outcome == "skipped" and current_result is None and self.config.target_selection is not None)
-
-    @staticmethod
-    def _build_inflight_row_progress_result(check: AuditCheck, row_id: str) -> RowEvaluationResult:
-        """Create an in-memory placeholder used only for row-start progress telemetry."""
-        return RowEvaluationResult(
-            row_id=row_id,
-            check_id=check.id,
-            status=RowStatus.NA,
-            score=0.0,
-            score_breakdown=ScoreBreakdown.compute(0.0),
-            reason="Đang chờ đánh giá theo tiến độ batch.",
-            finding="Kết quả sẽ được cập nhật khi batch chứa hàng này hoàn tất.",
-            evidence_anchors=[],
-            provenance=RowProvenance(
-                row_llm_mode="selective",
-                protocol=None,
-                model=None,
-                template_hash="",
-                rubric_version="1",
-                evidence_hash="",
-                evaluated_at=datetime.now(UTC),
-            ),
-        )
-
-    def _should_evaluate_row(
-        self,
-        row_index: int,
-        check_id: str,
-        row_id: str,
-        checkpoint: BatchCheckpoint,
-    ) -> bool:
-        """Determine if a row should be evaluated (FR-25.3, FR-25.4, FR-185).
-
-        Considers:
-        - Whether the row is scoped N/A by repo-type policy (FR-185)
-        - Whether the row is targeted
-        - Whether the row has prior checkpoint results
-        - Whether force_refresh_targets is enabled
-
-        Args:
-            row_index: 0-based row index.
-            check_id: The check ID for this row.
-            row_id: The unique row identifier.
-            checkpoint: Current checkpoint state.
-
-        Returns:
-            True if the row should be evaluated, False if it should be skipped
-            or checkpoint result should be used.
-        """
-        # FR-185: Skip rows marked N/A by repo-type scoping policy entirely.
-        # This prevents evidence retrieval, prompt construction, and LLM calls
-        # for structurally inapplicable rows (e.g., DB rows for frontend repos).
-        if row_index in self.config.scoped_na_indices:
-            return False
-
-        is_targeted = self._is_row_targeted(row_index, check_id)
-        has_checkpoint_result = row_id in checkpoint.results_by_row_id
-
-        if not is_targeted:
-            # Non-targeted rows: never evaluate, just preserve checkpoint if available
-            return False
-
-        # Targeted rows
-        if self.config.force_refresh_targets:
-            # FR-25.4: Re-evaluate targeted rows even if checkpointed
-            return True
-
-        # Only evaluate if no checkpoint result exists
-        return not has_checkpoint_result
-
-    async def process_all(
-        self,
-        thread_id: str,
-        project_profile: dict[str, Any] | None = None,
-        progress_callback: BatchProgressCallback | None = None,
-    ) -> list[RowEvaluationResult]:
-        """Process all rows in batches with checkpoint/resume (FR-3.3) and targeting (FR-25)."""
-        # FR-145.7: Reset resilience accumulator so each run starts fresh.
-        self._resilience_stats_accumulator = []
-        self.last_resilience_summary = None
-        total_rows = len(self._checks)
-
-        # Load checkpoint if resuming
-        checkpoint_path = self._get_checkpoint_path(thread_id)
-        checkpoint: BatchCheckpoint | None = None
-        if self.config.resume_from_checkpoint:
-            checkpoint = BatchCheckpoint.load(checkpoint_path)
-            if checkpoint:
-                # Validate checkpoint matches current template/evidence
-                current_evidence_hash = self.evaluator.evidence_hash or ""
-                if (
-                    checkpoint.template_hash != self.evaluator.template_hash
-                    or checkpoint.evidence_hash != current_evidence_hash
-                ):
-                    logger.warning(
-                        "checkpoint_hash_mismatch_starting_fresh",
-                        checkpoint_template=checkpoint.template_hash,
-                        current_template=self.evaluator.template_hash,
-                    )
-                    checkpoint = None
-                else:
-                    logger.info(
-                        "resuming_from_checkpoint",
-                        completed_rows=len(checkpoint.results_by_row_id),
-                        total_rows=total_rows,
-                    )
-
-        if not checkpoint:
-            checkpoint = BatchCheckpoint(
-                thread_id=thread_id,
-                total_rows=total_rows,
-                template_hash=self.evaluator.template_hash,
-                evidence_hash=self.evaluator.evidence_hash or "",
-                completed_batches=[],
-                results_by_row_id={},
-            )
-
-        # FR-25: Log targeting info if configured
-        if self.config.target_selection:
-            target_count = len(self.config.target_selection.row_indices)
-            logger.info(
-                "targeted_execution_mode",
-                target_type=self.config.target_selection.target_type.value,
-                target_count=target_count,
-                total_rows=total_rows,
-                force_refresh=self.config.force_refresh_targets,
-            )
-
-        # FR-3.3: Seed results from checkpoint for complete result set on resume
-        # FR-25.3: For non-targeted rows, preserve checkpoint state
-        all_results: list[RowEvaluationResult] = []
-        checkpointed_row_ids: set[str] = set()
-
-        for row_data in checkpoint.results_by_row_id.values():
-            try:
-                result = BatchCheckpoint.deserialize_row_result(row_data)
-                checkpointed_row_ids.add(result.row_id)
-                all_results.append(result)
-            except Exception as e:
-                logger.warning("failed_to_deserialize_checkpoint_row", error=str(e))
-
-        # Calculate batches
-        batch_count = (total_rows + self.config.batch_size - 1) // self.config.batch_size
-        batches_skipped = 0
-        batches_processed = 0
-
-        async def _emit_row_progress(
-            batch_idx: int,
-            start_idx: int,
-            end_idx: int,
-            batch_results_so_far: list[RowEvaluationResult],
-            success_count_so_far: int,
-            error_count_so_far: int,
-            skipped_count_so_far: int,
-            current_result: RowEvaluationResult | None,
-            outcome: str,
-        ) -> None:
-            # Non-targeted rows without checkpoint state return no concrete result.
-            # Suppress row-level telemetry for those no-op skips to avoid null-id noise.
-            if outcome == "skipped" and current_result is None:
-                return
-
-            preview_results = list(all_results)
-            if outcome != "started":
-                for result in batch_results_so_far:
-                    preview_results = [res for res in preview_results if res.row_id != result.row_id]
-                    preview_results.append(result)
-                    checkpoint.results_by_row_id[result.row_id] = BatchCheckpoint.serialize_row_result(result)
-
-                checkpoint.last_updated = datetime.now(UTC)
-                checkpoint.save(checkpoint_path)
-
-            completed_rows = len(preview_results)
-            target_row_count = (
-                len(self.config.target_selection.row_indices) if self.config.target_selection is not None else None
-            )
-            targeted_completed_rows = (
-                min(completed_rows, target_row_count) if target_row_count is not None else completed_rows
-            )
-            completion_scope = "targeted" if target_row_count is not None else "template"
-            completion_pct = round((completed_rows / total_rows) * 100.0 if total_rows > 0 else 100.0, 2)
-            targeted_completion_pct = round(
-                (targeted_completed_rows / target_row_count) * 100.0 if target_row_count else completion_pct,
-                2,
-            )
-            progress = BatchProgress(
-                batch_index=batch_idx,
-                batch_count=batch_count,
-                start_row=start_idx,
-                end_row=end_idx,
-                total_rows=total_rows,
-                completed_rows=completed_rows,
-                processed_batches=batches_processed,
-                skipped_batches=batches_skipped,
-                success_count=success_count_so_far,
-                error_count=error_count_so_far,
-                skipped_count=skipped_count_so_far,
-                duration_ms=0,
-                timed_out=False,
-                emission_kind="row",
-                current_row_id=getattr(current_result, "row_id", None),
-                current_check_id=getattr(current_result, "check_id", None),
-            )
-            self.last_progress = progress
-            progress_context = dict(self.config.run_context or {})
-            if current_result is not None:
-                provenance = current_result.provenance
-                retrieval_trace = (
-                    current_result.retrieval_trace if isinstance(current_result.retrieval_trace, dict) else {}
-                )
-                row_distribution_profile = str(retrieval_trace.get("row_distribution_profile") or "").strip() or None
-                original_provider = (
-                    str(getattr(provenance, "original_provider", "") or "").strip() if provenance is not None else ""
-                ) or None
-                final_provider = (
-                    str(getattr(provenance, "final_provider", "") or "").strip() if provenance is not None else ""
-                ) or None
-                executed_profile = row_distribution_profile or final_provider or original_provider
-                if executed_profile:
-                    progress_context["executed_profile"] = executed_profile
-                    progress_context["active_profile"] = executed_profile
-                if original_provider:
-                    progress_context["origin_profile"] = original_provider
-                if row_distribution_profile:
-                    progress_context["row_distribution_profile"] = row_distribution_profile
-            logger.info(
-                "row_progress",
-                batch=f"{progress.batch_index + 1}/{progress.batch_count}",
-                completed_rows=progress.completed_rows,
-                total_rows=progress.total_rows,
-                completion_pct=completion_pct,
-                completion_scope=completion_scope,
-                targeted_completed_rows=targeted_completed_rows if target_row_count is not None else None,
-                targeted_total_rows=target_row_count,
-                targeted_completion_pct=targeted_completion_pct if target_row_count is not None else None,
-                processed_batches=progress.processed_batches,
-                skipped_batches=progress.skipped_batches,
-                row_id=progress.current_row_id,
-                check_id=progress.current_check_id,
-                row_outcome=outcome,
-                thread_id=thread_id,
-                **progress_context,
-            )
-            if progress_callback is not None:
-                try:
-                    maybe_awaitable = progress_callback(progress, list(preview_results))
-                    if inspect.isawaitable(maybe_awaitable):
-                        await maybe_awaitable
-                except Exception as e:
-                    logger.warning(
-                        "row_progress_callback_failed",
-                        batch=progress.batch_index + 1,
-                        row_id=progress.current_row_id,
-                        error=str(e),
-                    )
-
-        for batch_idx in range(batch_count):
-            start_idx = batch_idx * self.config.batch_size
-            end_idx = min(start_idx + self.config.batch_size, total_rows)
-
-            # FR-25.2: Skip whole batches when they contain no targeted rows
-            if not self._batch_has_targeted_rows(start_idx, end_idx):
-                logger.debug(
-                    "skipping_batch_no_targets",
-                    batch=batch_idx,
-                    rows=f"{start_idx + 1}-{end_idx}",
-                )
-                batches_skipped += 1
-                continue
-
-            # Skip completed batches (only if not in targeted mode with force_refresh)
-            # In targeted mode with force_refresh, we need to re-process targeted rows
-            if (batch_idx in checkpoint.completed_batches and self.config.target_selection is None) or not self.config.force_refresh_targets:
-                    logger.debug("skipping_completed_batch", batch=batch_idx)
-                    continue
-                # In force_refresh mode, we still process the batch but only targeted rows
-
-            # FR-3: Progress logging
-            logger.info(
-                "processing_batch",
-                batch=f"{batch_idx + 1}/{batch_count}",
-                rows=f"{start_idx + 1}-{end_idx}",
-                targeted=self.config.target_selection is not None,
-            )
-
-            # Process batch with timeout (FR-3.2)
-            try:
-                batch_result = await self._process_batch(
-                    batch_idx=batch_idx,
-                    start_idx=start_idx,
-                    end_idx=end_idx,
-                    project_profile=project_profile,
-                    checkpoint=checkpoint,
-                    thread_id=thread_id,
-                    row_progress_callback=_emit_row_progress,
-                )
-            except BatchBudgetExceededError as exc:
-                for r in exc.partial_results:
-                    all_results = [res for res in all_results if res.row_id != r.row_id]
-                    all_results.append(r)
-                    checkpoint.results_by_row_id[r.row_id] = BatchCheckpoint.serialize_row_result(r)
-
-                checkpoint.last_updated = datetime.now(UTC)
-                checkpoint.save(checkpoint_path)
-                exc.context["checkpoint_path"] = str(checkpoint_path)
-                exc.context["completed_rows_persisted"] = len(all_results)
-                exc.partial_results = list(all_results)
-                logger.exception("batch_budget_exceeded", **exc.context)
-                raise
-
-            # FR-25.3: Merge results - update targeted rows, preserve non-targeted
-            for r in batch_result.results:
-                # Remove old result if exists (for force_refresh case)
-                all_results = [res for res in all_results if res.row_id != r.row_id]
-                all_results.append(r)
-                # Update checkpoint
-                checkpoint.results_by_row_id[r.row_id] = BatchCheckpoint.serialize_row_result(r)
-
-            # FR-3.3: Update checkpoint
-            if batch_idx not in checkpoint.completed_batches:
-                checkpoint.completed_batches.append(batch_idx)
-            checkpoint.last_updated = datetime.now(UTC)
-            checkpoint.save(checkpoint_path)
-
-            batches_processed += 1
-
-            progress = BatchProgress(
-                batch_index=batch_idx,
-                batch_count=batch_count,
-                start_row=start_idx,
-                end_row=end_idx,
-                total_rows=total_rows,
-                completed_rows=len(all_results),
-                processed_batches=batches_processed,
-                skipped_batches=batches_skipped,
-                success_count=batch_result.success_count,
-                error_count=batch_result.error_count,
-                skipped_count=batch_result.skipped_count,
-                duration_ms=batch_result.duration_ms,
-                timed_out=batch_result.timed_out,
-            )
-            self.last_progress = progress
-            target_row_count = (
-                len(self.config.target_selection.row_indices) if self.config.target_selection is not None else None
-            )
-            targeted_completed_rows = (
-                min(progress.completed_rows, target_row_count)
-                if target_row_count is not None
-                else progress.completed_rows
-            )
-            completion_scope = "targeted" if target_row_count is not None else "template"
-            targeted_completion_pct = round(
-                (targeted_completed_rows / target_row_count) * 100.0 if target_row_count else 100.0,
-                2,
-            )
-            logger.info(
-                "batch_progress",
-                batch=f"{progress.batch_index + 1}/{progress.batch_count}",
-                completed_rows=progress.completed_rows,
-                total_rows=progress.total_rows,
-                completion_pct=round(
-                    (progress.completed_rows / progress.total_rows) * 100.0 if progress.total_rows > 0 else 100.0,
-                    2,
-                ),
-                completion_scope=completion_scope,
-                targeted_completed_rows=targeted_completed_rows if target_row_count is not None else None,
-                targeted_total_rows=target_row_count,
-                targeted_completion_pct=targeted_completion_pct if target_row_count is not None else None,
-                processed_batches=progress.processed_batches,
-                skipped_batches=progress.skipped_batches,
-                thread_id=thread_id,
-                **self.config.run_context,
-            )
-
-            if progress_callback is not None:
-                try:
-                    maybe_awaitable = progress_callback(progress, list(all_results))
-                    if inspect.isawaitable(maybe_awaitable):
-                        await maybe_awaitable
-                except Exception as e:
-                    logger.warning(
-                        "batch_progress_callback_failed",
-                        batch=progress.batch_index + 1,
-                        error=str(e),
-                    )
-
-            logger.info(
-                "batch_complete",
-                batch=batch_idx + 1,
-                success=batch_result.success_count,
-                errors=batch_result.error_count,
-                skipped=batch_result.skipped_count,
-                duration_ms=batch_result.duration_ms,
-                timed_out=batch_result.timed_out,
-            )
-
-        # FR-25: Log summary
-        if self.config.target_selection:
-            target_row_count = len(self.config.target_selection.row_indices)
-            logger.info(
-                "targeted_execution_complete",
-                batches_processed=batches_processed,
-                batches_skipped=batches_skipped,
-                total_results=len(all_results),
-                completion_scope="targeted",
-                targeted_completed_rows=len(all_results),
-                targeted_total_rows=target_row_count,
-                targeted_completion_pct=round(
-                    (len(all_results) / target_row_count) * 100.0 if target_row_count else 100.0, 2
-                ),
-                thread_id=thread_id,
-                **self.config.run_context,
-            )
-        self.last_execution_summary = {
-            "processed_batches": batches_processed,
-            "skipped_batches": batches_skipped,
-            "total_results": len(all_results),
-            "batch_count": batch_count,
-        }
-
-        # Phase 166: Accumulate prompt-level retry/failover telemetry from row results.
-        _prompt_stats = _ResilienceRetryStats()
-        for row_result in all_results:
-            tel = self._extract_prompt_level_telemetry(row_result)
-            _prompt_stats.prompt_retry_attempts += tel["prompt_retry_attempts"]
-            _prompt_stats.prompt_retry_recoveries += tel["prompt_retry_recoveries"]
-            _prompt_stats.prompt_failover_attempts += tel["prompt_failover_attempts"]
-            _prompt_stats.prompt_failover_recoveries += tel["prompt_failover_recoveries"]
-            _prompt_stats.prompt_failover_exhausted += tel["prompt_failover_exhausted"]
-            _prompt_stats.synthesis_fallback_count += tel["synthesis_fallback_count"]
-        self._resilience_stats_accumulator.append(_prompt_stats)
-
-        # FR-145.7: Build and emit full resilience_summary at batch completion.
-        # Always build, regardless of whether retries ran — callers rely on the dict.
-        final_error_count = sum(1 for r in all_results if r.status == RowStatus.ERROR)
-        # error_rows_before_retry: sum of "before" snapshots across all batch stats.
-        error_rows_before_retry = sum(s.error_rows_before_retry for s in self._resilience_stats_accumulator)
-        retry_attempts = sum(s.retry_attempts for s in self._resilience_stats_accumulator)
-        retry_successes = sum(s.retry_successes for s in self._resilience_stats_accumulator)
-        retry_budget_used = sum(s.budget_used for s in self._resilience_stats_accumulator)
-        retry_budget_total = sum(s.budget_total for s in self._resilience_stats_accumulator)
-        failover_hops_total = sum(s.failover_hops for s in self._resilience_stats_accumulator)
-        quota_exhausted_rows = sum(s.quota_exhausted_rows for s in self._resilience_stats_accumulator)
-        quota_wait_attempted_rows = sum(s.quota_wait_attempted_rows for s in self._resilience_stats_accumulator)
-        quota_wait_recovered_rows = sum(s.quota_wait_recovered_rows for s in self._resilience_stats_accumulator)
-        # Phase 166: Aggregate prompt-level retry/failover telemetry.
-        prompt_retry_attempts = sum(s.prompt_retry_attempts for s in self._resilience_stats_accumulator)
-        prompt_retry_recoveries = sum(s.prompt_retry_recoveries for s in self._resilience_stats_accumulator)
-        prompt_failover_attempts = sum(s.prompt_failover_attempts for s in self._resilience_stats_accumulator)
-        prompt_failover_recoveries = sum(s.prompt_failover_recoveries for s in self._resilience_stats_accumulator)
-        prompt_failover_exhausted = sum(s.prompt_failover_exhausted for s in self._resilience_stats_accumulator)
-        synthesis_fallback_count = sum(s.synthesis_fallback_count for s in self._resilience_stats_accumulator)
-        # Deduplicate providers used across batches.
-        providers_used: list[str] = []
-        for s in self._resilience_stats_accumulator:
-            for p in s.providers_used:
-                if p not in providers_used:
-                    providers_used.append(p)
-        # Build provider_health_summary from shared health memory if available.
-        health_memory: ProviderHealthMemory | None = getattr(self, "_batch_health_memory", None)
-        provider_health_summary: dict[str, dict[str, Any]] = {}
-        if health_memory is not None:
-            all_providers = (
-                set(health_memory.consecutive_failures)
-                | set(health_memory.total_retries)
-                | set(health_memory.cooldown_until)
-                | set(health_memory.states)
-            )
-            for provider in sorted(all_providers):
-                state = health_memory.get_state(provider)
-                provider_health_summary[provider] = {
-                    "failures": health_memory.consecutive_failures.get(provider, 0),
-                    "cooldowns": 1 if health_memory.is_in_cooldown(provider) else 0,
-                    "retries": health_memory.total_retries.get(provider, 0),
-                    "quota_status": state.quota_status.value,
-                    "last_reason_code": state.last_reason_code,
-                    "retry_after_seconds": state.rate_limit.retry_after_seconds,
-                    "remaining_requests": state.rate_limit.remaining_requests,
-                    "remaining_tokens": state.rate_limit.remaining_tokens,
-                }
-        self.last_resilience_summary = {
-            "total_rows": total_rows,
-            "error_rows_before_retry": error_rows_before_retry,
-            "error_rows_after_retry": final_error_count,
-            "retry_attempts": retry_attempts,
-            "retry_successes": retry_successes,
-            "retry_budget_used": retry_budget_used,
-            "retry_budget_total": retry_budget_total,
-            "failover_hops_total": failover_hops_total,
-            "providers_used": providers_used,
-            "provider_health_summary": provider_health_summary,
-            # Phase 146 quota-exhaustion telemetry (AC-146.7.3)
-            "quota_exhausted_rows": quota_exhausted_rows,
-            "quota_wait_attempted_rows": quota_wait_attempted_rows,
-            "quota_wait_recovered_rows": quota_wait_recovered_rows,
-            # Phase 166: Prompt-level retry/failover telemetry
-            "prompt_retry_attempts": prompt_retry_attempts,
-            "prompt_retry_recoveries": prompt_retry_recoveries,
-            "prompt_failover_attempts": prompt_failover_attempts,
-            "prompt_failover_recoveries": prompt_failover_recoveries,
-            "prompt_failover_exhausted": prompt_failover_exhausted,
-            "synthesis_fallback_count": synthesis_fallback_count,
-        }
-        logger.info(
-            "resilience_summary",
-            **self.last_resilience_summary,
-            thread_id=thread_id,
-        )
-
-        return all_results
-
-    async def _process_batch(
-        self,
-        batch_idx: int,
-        start_idx: int,
-        end_idx: int,
-        project_profile: dict[str, Any] | None,
-        checkpoint: BatchCheckpoint,
-        thread_id: str,
-        row_progress_callback: RowProgressCallback | None = None,
-    ) -> BatchResult:
-        """Process a single batch of rows with per-row timeout (FR-3.2) and targeting (FR-25).
-
-        In mixed batches (some targeted, some not), only targeted rows are evaluated.
-        Non-targeted rows preserve their checkpoint state (FR-25.3).
-
-        When row_concurrency > 1 (FR-188), rows are dispatched in parallel via
-        asyncio.gather with a semaphore-bounded concurrency limit.
-        """
-        self._batch_match_cache = {}
-        self._batch_match_inflight = {}
-        inline_batch_stats = _ResilienceRetryStats()
-        # FR-145.4 (AC-145.4.1): Create one shared ProviderHealthMemory per batch run.
-        # Stored as an instance attribute so _process_single_row (parallel path) and
-        # the serial loop below both share the same instance without extra arg threading.
-        self._batch_health_memory = ProviderHealthMemory()
-        await self._seed_batch_health_memory_from_preflight()
-        # FR-145.1 (AC-145.1.1): Resolve default failover profiles for this batch.
-        # When the operator has not explicitly configured row_failover_profiles, auto-derive
-        # from all configured runtime profiles minus the active profile.
-        import os as _os
-
-        _failover_env = _os.getenv("VDS_AUDIT_LLM__ROW_FAILOVER_PROFILES")
-        if _failover_env is None:
-            _active = str(_os.getenv("VDS_AUDIT_ACTIVE_PROFILE") or "").strip() or None
-            self._batch_failover_profiles: list[str] = resolve_default_failover_profiles(_active)
-            if self._batch_failover_profiles:
-                logger.info(
-                    "provider_failover_auto_enabled",
-                    active_profile=_active,
-                    failover_profiles=self._batch_failover_profiles,
-                    source="auto_derived",
-                )
-        else:
-            import json as _json
-
-            try:
-                _raw = _json.loads(_failover_env)
-                self._batch_failover_profiles = (
-                    [str(p).strip() for p in _raw if str(p).strip()] if isinstance(_raw, list) else []
-                )
-            except Exception:
-                self._batch_failover_profiles = [p.strip() for p in _failover_env.split(",") if p.strip()]
-
-        # FR-188: Dispatch to parallel path when concurrency > 1
-        if self.config.row_concurrency > 1:
-            return await self._process_batch_parallel(
-                batch_idx=batch_idx,
-                start_idx=start_idx,
-                end_idx=end_idx,
-                project_profile=project_profile,
-                checkpoint=checkpoint,
-                thread_id=thread_id,
-                row_progress_callback=row_progress_callback,
-                batch_stats=inline_batch_stats,
-            )
-        import time
-
-        start_time = time.monotonic()
-        results: list[RowEvaluationResult] = []
-        success_count = 0
-        error_count = 0
-        skipped_count = 0
-        timed_out = False
-        next_row_index = start_idx
-
-        try:
-            effective_batch_timeout_ms = self.effective_batch_timeout_ms()
-            async with asyncio.timeout(effective_batch_timeout_ms / 1000):
-                for idx in range(start_idx, end_idx):
-                    next_row_index = idx
-                    section_id, check = self._checks[idx]
-                    row_id = f"{check.id}:row_{idx}"
-
-                    # FR-25.3, FR-25.4: Check if row should be evaluated
-                    if not self._should_evaluate_row(idx, check.id, row_id, checkpoint):
-                        # Preserve checkpoint state for non-targeted rows
-                        preserved_result: RowEvaluationResult | None = None
-                        if row_id in checkpoint.results_by_row_id:
-                            try:
-                                preserved_result = BatchCheckpoint.deserialize_row_result(
-                                    checkpoint.results_by_row_id[row_id]
-                                )
-                                results.append(preserved_result)
-                                logger.debug(
-                                    "preserving_checkpoint_result",
-                                    row_id=row_id,
-                                    row_index=idx,
-                                )
-                            except Exception as e:
-                                logger.warning(
-                                    "failed_to_preserve_checkpoint_result",
-                                    row_id=row_id,
-                                    error=str(e),
-                                )
-                        skipped_count += 1
-                        if row_progress_callback is not None and self._should_emit_row_progress_callback(
-                            preserved_result,
-                            "skipped",
-                        ):
-                            maybe_awaitable = row_progress_callback(
-                                batch_idx,
-                                start_idx,
-                                end_idx,
-                                list(results),
-                                success_count,
-                                error_count,
-                                skipped_count,
-                                preserved_result,
-                                "skipped",
-                            )
-                            if inspect.isawaitable(maybe_awaitable):
-                                await maybe_awaitable
-                        next_row_index = idx + 1
-                        continue
-
-                    if row_progress_callback is not None:
-                        inflight_result = self._build_inflight_row_progress_result(check, row_id)
-                        maybe_awaitable = row_progress_callback(
-                            batch_idx,
-                            start_idx,
-                            end_idx,
-                            list(results),
-                            success_count,
-                            error_count,
-                            skipped_count,
-                            inflight_result,
-                            "started",
-                        )
-                        if inspect.isawaitable(maybe_awaitable):
-                            await maybe_awaitable
-
-                    match_requirement_text = self._build_match_requirement_text(check)
-                    app_config_only = self._row_requires_app_config_only(check)
-
-                    # App-config-only rows must bypass generic matcher retrieval entirely.
-                    evidence_context = await self._get_or_match_evidence_context(
-                        row_id=row_id,
-                        requirement_text=match_requirement_text,
-                        section_id=section_id,
-                        app_config_only=app_config_only,
-                    )
-
-                    formatted_evidence = "" if app_config_only else evidence_context.format_for_llm()
-                    evidence_refs = evidence_context.evidence_refs  # FR-2.4
-
-                    try:
-                        result = await self._evaluate_row_with_timeout_mode(
-                            check=check,
-                            row_id=row_id,
-                            row_index=idx,
-                            evidence_context=formatted_evidence,
-                            evidence_refs=evidence_refs,
-                            project_profile=project_profile,
-                            thread_id=thread_id,
-                            matcher_retrieval_trace=getattr(evidence_context, "retrieval_trace", None),
-                            provider_health_memory=self._batch_health_memory,
-                        )
-                    except BudgetExceededError as exc:
-                        usage = exc.usage
-                        context = {
-                            "kind": "strict_budget_exceeded",
-                            "audit_error_key": "LLM_BUDGET_EXCEEDED",
-                            "audit_error_code": AUDIT_ERROR_CODES["LLM_BUDGET_EXCEEDED"].code,
-                            "batch_index": batch_idx + 1,
-                            "row_index": idx + 1,
-                            "row_id": row_id,
-                            "check_id": check.id,
-                            "scope": usage.scope.value,
-                            "scope_id": usage.scope_id,
-                            "used_dollars": usage.used_dollars,
-                            "limit_dollars": usage.limit_dollars,
-                            "status": usage.status.value,
-                            "error_context": getattr(exc, "error_context", None),
-                        }
-                        raise BatchBudgetExceededError(
-                            message=str(exc),
-                            context=context,
-                            partial_results=list(results),
-                            success_count=success_count,
-                            error_count=error_count,
-                            skipped_count=skipped_count,
-                        ) from exc
-                    finally:
-                        self._clear_row_runtime_progress(row_id)
-
-                    matcher_trace = _sanitize_retrieval_trace(getattr(evidence_context, "retrieval_trace", None))
-                    row_trace = _sanitize_retrieval_trace(result.retrieval_trace)
-                    merged_trace = _merge_retrieval_trace_payloads(matcher_trace, row_trace)
-                    if isinstance(merged_trace, dict):
-                        # Preserve matcher retrieval context (docs/code candidates, backend mode)
-                        # while allowing row-evaluator trace keys to take precedence.
-                        result.retrieval_trace = _sanitize_retrieval_trace(merged_trace)
-
-                    # TSK-145.13/14: Timeout-aware failover — attempt ONE failover
-                    # evaluation for STALL/NO_PROGRESS/CHURN_DETECTED timeouts before
-                    # finalising ERROR. TIMEOUT_WITH_GROUNDING rows are excluded.
-                    if result.status == RowStatus.ERROR:
-                        timeout_kind_in_trace = str((result.retrieval_trace or {}).get("timeout_kind") or "").strip()
-                        if self._is_timeout_failoverable(timeout_kind_in_trace) and self._batch_failover_profiles:
-                            result = await self._attempt_timeout_failover(
-                                timeout_result=result,
-                                check=check,
-                                row_index=idx,
-                                evidence_context=formatted_evidence,
-                                evidence_refs=evidence_refs,
-                                project_profile=project_profile,
-                                health_memory=self._batch_health_memory,
-                                failover_profiles=self._batch_failover_profiles,
-                                stats=inline_batch_stats,
-                            )
-
-                    # FR-7: Validate grounding if validator configured
-                    if self.grounding_validator:
-                        result = self.grounding_validator.validate(
-                            result,
-                            allowed_refs=evidence_refs,
-                            evidence_context=formatted_evidence,
-                        )
-
-                    results.append(result)
-
-                    if result.status == RowStatus.ERROR:
-                        error_count += 1
-                    else:
-                        success_count += 1
-                    if row_progress_callback is not None:
-                        maybe_awaitable = row_progress_callback(
-                            batch_idx,
-                            start_idx,
-                            end_idx,
-                            list(results),
-                            success_count,
-                            error_count,
-                            skipped_count,
-                            result,
-                            "error" if result.status == RowStatus.ERROR else "success",
-                        )
-                        if inspect.isawaitable(maybe_awaitable):
-                            await maybe_awaitable
-                    next_row_index = idx + 1
-
-        except TimeoutError:
-            logger.warning(
-                "batch_timeout",
-                batch=batch_idx,
-                completed=len(results),
-                expected=end_idx - start_idx,
-            )
-            timed_out = True
-
-            # Mark remaining rows as ERROR (FR-3.2: graceful degradation)
-            for idx in range(next_row_index, end_idx):
-                section_id, check = self._checks[idx]
-                row_id = f"{check.id}:row_{idx}"
-
-                # FR-25: Only create error for targeted rows, preserve checkpoint for others
-                if self._should_evaluate_row(idx, check.id, row_id, checkpoint):
-                    progress = self._normalize_timeout_progress_snapshot(self._get_row_runtime_progress(row_id))
-                    degraded_timeout_finalize = self._should_degrade_timeout_finalize(progress)
-                    timeout_result = self._create_timeout_result(
-                        row_id,
-                        check.id,
-                        "Batch timeout exceeded",
-                        degraded_timeout_finalize=degraded_timeout_finalize,
-                        progress=progress,
-                    )
-                    timeout_trace = {
-                        "timeout_retry_attempts": 0,
-                        "timeout_progress_snapshot": progress,
-                        "timeout_extended": False,
-                        "timeout_degraded_finalize": degraded_timeout_finalize,
-                        "timeout_kind": self._derive_timeout_kind(
-                            progress=progress,
-                            degraded_timeout_finalize=degraded_timeout_finalize,
-                            batch_timeout_exceeded=True,
-                        ),
-                        "timeout_terminal_status": timeout_result.status.value,
-                        "batch_timeout_exceeded": True,
-                    }
-                    self._attach_timeout_progress_telemetry(
-                        timeout_trace=timeout_trace,
-                        progress=progress,
-                        row_id=row_id,
-                        check_id=check.id,
-                    )
-                    timeout_result.retrieval_trace = _sanitize_retrieval_trace(
-                        self._sanitize_timeout_trace_for_app_config_only(
-                            timeout_trace,
-                            progress=progress,
-                            matcher_retrieval_trace=None,
-                        )
-                    )
-                    timeout_result = RowEvaluator._sanitize_app_config_only_result(timeout_result)
-                    timeout_result = RowEvaluator._salvage_app_config_only_timeout_result(timeout_result)
-                    results.append(timeout_result)
-                    error_count += 1
-                elif row_id in checkpoint.results_by_row_id:
-                    try:
-                        preserved_result = BatchCheckpoint.deserialize_row_result(checkpoint.results_by_row_id[row_id])
-                        results.append(preserved_result)
-                    except Exception:
-                        pass
-                    skipped_count += 1
-
-        duration_ms = int((time.monotonic() - start_time) * 1000)
-
-        # FR-145.2: Post-pass ERROR retry sweep.
-        # Only runs when batch_error_retry_limit > 0 and there are ERROR rows.
-        # Requires _batch_health_memory and _batch_failover_profiles to be set
-        # (threaded by the batch loop — see TSK-145.1 / TSK-145.3).
-        if not timed_out and error_count > 0 and self.config.batch_error_retry_limit > 0:
-            health_memory: ProviderHealthMemory | None = getattr(self, "_batch_health_memory", None)
-            failover_profiles: list[str] = list(getattr(self, "_batch_failover_profiles", None) or [])
-            if health_memory is not None and failover_profiles:
-                budget = ResilienceBudget(
-                    total_rows=len(results),
-                    error_count=error_count,
-                )
-                retried = await self._retry_error_rows(
-                    results=results,
-                    health_memory=health_memory,
-                    budget=budget,
-                    failover_profiles=failover_profiles,
-                    project_profile=project_profile,
-                    stats=inline_batch_stats,
-                )
-                # Recount after retry pass.
-                new_error_count = sum(1 for r in retried if r.status == RowStatus.ERROR)
-                new_success_count = sum(
-                    1
-                    for r in retried
-                    if r.status != RowStatus.ERROR
-                    and r.row_id in {orig.row_id for orig in results if orig.status == RowStatus.ERROR}
-                )
-                if budget.calls_used > 0:
-                    logger.info(
-                        "resilience_retry_sweep_complete",
-                        batch_index=batch_idx,
-                        errors_before=error_count,
-                        errors_after=new_error_count,
-                        retries_attempted=budget.calls_used,
-                        budget_total=budget.total,
-                    )
-                results = retried
-                error_count = new_error_count
-                success_count += new_success_count
-
-                # TSK-145.9: Low-confidence re-evaluation sweep.
-                # Shares the same budget instance — any budget consumed by ERROR
-                # retries above reduces what is available here.
-                reevaled = await self._reevaluate_low_confidence_rows(
-                    results=results,
-                    health_memory=health_memory,
-                    budget=budget,
-                    failover_profiles=failover_profiles,
-                    project_profile=project_profile,
-                )
-                reeval_improved = sum(
-                    1
-                    for orig, new in zip(results, reevaled, strict=False)
-                    if (new.score_1_5 or 0) > (orig.score_1_5 or 0)
-                )
-                if reeval_improved > 0:
-                    logger.info(
-                        "low_confidence_reeval_sweep_complete",
-                        batch_index=batch_idx,
-                        rows_improved=reeval_improved,
-                        budget_calls_used=budget.calls_used,
-                        budget_total=budget.total,
-                    )
-                results = reevaled
-
-        self._resilience_stats_accumulator.append(inline_batch_stats)
-        return BatchResult(
-            batch_index=batch_idx,
-            start_row=start_idx,
-            end_row=end_idx,
-            results=results,
-            duration_ms=duration_ms,
-            success_count=success_count,
-            error_count=error_count,
-            timed_out=timed_out,
-            skipped_count=skipped_count,
-        )
-
-    async def _process_single_row(
-        self,
-        idx: int,
-        batch_idx: int,
-        project_profile: dict[str, Any] | None,
-        checkpoint: BatchCheckpoint,
-        thread_id: str,
-        semaphore: asyncio.Semaphore,
-        checkpoint_lock: asyncio.Lock,
-        batch_stats: _ResilienceRetryStats | None = None,
-    ) -> tuple[RowEvaluationResult | None, str]:
-        """Process a single row under semaphore guard (FR-188).
-
-        Returns (result, outcome) where outcome is 'success', 'error', or 'skipped'.
-        Budget exceptions propagate to the caller.
-        """
-        section_id, check = self._checks[idx]
-        row_id = f"{check.id}:row_{idx}"
-
-        # FR-25.3, FR-25.4: Check if row should be processed
-        if not self._should_evaluate_row(idx, check.id, row_id, checkpoint):
-            async with checkpoint_lock:
-                if row_id in checkpoint.results_by_row_id:
-                    try:
-                        preserved_result = BatchCheckpoint.deserialize_row_result(checkpoint.results_by_row_id[row_id])
-                        logger.debug(
-                            "preserving_checkpoint_result",
-                            row_id=row_id,
-                            row_index=idx,
-                        )
-                        return preserved_result, "skipped"
-                    except Exception as e:
-                        logger.warning(
-                            "failed_to_preserve_checkpoint_result",
-                            row_id=row_id,
-                            error=str(e),
-                        )
-            return None, "skipped"
-
-        async with semaphore:
-            match_requirement_text = self._build_match_requirement_text(check)
-            app_config_only = self._row_requires_app_config_only(check)
-
-            evidence_context = await self._get_or_match_evidence_context(
-                row_id=row_id,
-                requirement_text=match_requirement_text,
-                section_id=section_id,
-                app_config_only=app_config_only,
-            )
-
-            formatted_evidence = "" if app_config_only else evidence_context.format_for_llm()
-            evidence_refs = evidence_context.evidence_refs
-
-            result: RowEvaluationResult
-            try:
-                result = await self._evaluate_row_with_timeout_mode(
-                    check=check,
-                    row_id=row_id,
-                    row_index=idx,
-                    evidence_context=formatted_evidence,
-                    evidence_refs=evidence_refs,
-                    project_profile=project_profile,
-                    thread_id=thread_id,
-                    matcher_retrieval_trace=getattr(evidence_context, "retrieval_trace", None),
-                    provider_health_memory=getattr(self, "_batch_health_memory", None),
-                )
-            except BudgetExceededError:
-                raise
-            finally:
-                self._clear_row_runtime_progress(row_id)
-
-            matcher_trace = _sanitize_retrieval_trace(getattr(evidence_context, "retrieval_trace", None))
-            row_trace = _sanitize_retrieval_trace(result.retrieval_trace)
-            merged_trace = _merge_retrieval_trace_payloads(matcher_trace, row_trace)
-            if isinstance(merged_trace, dict):
-                result.retrieval_trace = _sanitize_retrieval_trace(merged_trace)
-                result = RowEvaluator._sanitize_app_config_only_result(result)
-
-            # TSK-145.13/14: Timeout-aware failover — attempt ONE failover
-            # evaluation for STALL/NO_PROGRESS timeouts before finalising ERROR.
-            # TIMEOUT_WITH_GROUNDING rows are excluded (already have grounding).
-            if result.status == RowStatus.ERROR:
-                timeout_kind_in_trace = str((result.retrieval_trace or {}).get("timeout_kind") or "").strip()
-                if self._is_timeout_failoverable(timeout_kind_in_trace):
-                    health_memory_for_failover: ProviderHealthMemory | None = getattr(
-                        self, "_batch_health_memory", None
-                    )
-                    failover_profiles_for_timeout: list[str] = list(
-                        getattr(self, "_batch_failover_profiles", None) or []
-                    )
-                    if health_memory_for_failover is not None and failover_profiles_for_timeout:
-                        result = await self._attempt_timeout_failover(
-                            timeout_result=result,
-                            check=check,
-                            row_index=idx,
-                            evidence_context=formatted_evidence,
-                            evidence_refs=evidence_refs,
-                            project_profile=project_profile,
-                            health_memory=health_memory_for_failover,
-                            failover_profiles=failover_profiles_for_timeout,
-                        )
-
-            if self.grounding_validator:
-                result = self.grounding_validator.validate(
-                    result,
-                    allowed_refs=evidence_refs,
-                    evidence_context=formatted_evidence,
-                )
-
-            outcome = "error" if result.status == RowStatus.ERROR else "success"
-            return result, outcome
-
-    async def _evaluate_row_with_timeout_mode(
-        self,
-        *,
-        check: AuditCheck,
-        row_id: str,
-        row_index: int,
-        evidence_context: str,
-        evidence_refs: list[str],
-        project_profile: dict[str, Any] | None,
-        thread_id: str,
-        matcher_retrieval_trace: dict[str, Any] | None,
-        provider_health_memory: Any | None = None,
-    ) -> RowEvaluationResult:
-        if self._lease_timeout_mode_enabled():
-            return await self._evaluate_row_with_lease_timeout(
-                check=check,
-                row_id=row_id,
-                row_index=row_index,
-                evidence_context=evidence_context,
-                evidence_refs=evidence_refs,
-                project_profile=project_profile,
-                thread_id=thread_id,
-                matcher_retrieval_trace=matcher_retrieval_trace,
-                provider_health_memory=provider_health_memory,
-            )
-        return await self._evaluate_row_with_legacy_timeout(
-            check=check,
-            row_id=row_id,
-            row_index=row_index,
-            evidence_context=evidence_context,
-            evidence_refs=evidence_refs,
-            project_profile=project_profile,
-            thread_id=thread_id,
-            matcher_retrieval_trace=matcher_retrieval_trace,
-            provider_health_memory=provider_health_memory,
-        )
-
-    def _build_row_assigned_evaluator(self, assigned_profile: str) -> RowEvaluator:
-        runtime_context = self.config.run_context if isinstance(self.config.run_context, dict) else {}
-        cloned_config = copy.deepcopy(self.evaluator.config)
-        cloned_runtime_context = dict(getattr(cloned_config, "runtime_context", {}) or {})
-        source_llm = getattr(self.evaluator, "config", None)
-        source_llm = getattr(source_llm, "llm", None) if source_llm is not None else None
-        inherited_llm = inherit_runtime_llm_policy(assigned_profile, source_llm=source_llm)
-        cloned_runtime_context["repo_row_distribution_enabled"] = True
-        cloned_runtime_context["repo_distribution_profiles"] = list(self._repo_row_distribution_profiles())
-        cloned_runtime_context["row_distribution_profile"] = assigned_profile
-        cloned_runtime_context["active_profile"] = assigned_profile
-        cloned_runtime_context["repo_primary_profile"] = (
-            str(runtime_context.get("repo_primary_profile") or runtime_context.get("active_profile") or "").strip()
-            or assigned_profile
-        )
-        cloned_config.runtime_context = cloned_runtime_context
-        cloned_config.protocol = (
-            str(
-                getattr(getattr(inherited_llm, "protocol", None), "value", getattr(inherited_llm, "protocol", None))
-                or cloned_config.protocol
-                or ""
-            )
-            or None
-        )
-        cloned_config.model = str(getattr(inherited_llm, "model_standard", "") or cloned_config.model or "") or None
-        cloned_config.llm = inherited_llm
-        return RowEvaluator(
-            config=cloned_config,
-            template=self.template,
-            evidence_bundle=self.evaluator.evidence_bundle,
-        )
-
-    @staticmethod
-    def _apply_assigned_profile_provenance(
-        result: RowEvaluationResult,
-        *,
-        assigned_profile: str,
-    ) -> RowEvaluationResult:
-        provenance = (
-            result.provenance.model_copy(deep=True)
-            if result.provenance is not None
-            else RowProvenance(
-                row_llm_mode="unknown",
-                template_hash="unknown",
-                evidence_hash="unknown",
-            )
-        )
-        if not provenance.original_provider:
-            provenance.original_provider = assigned_profile
-        if not provenance.final_provider:
-            provenance.final_provider = assigned_profile
-        if not provenance.failover_chain:
-            provenance.failover_chain = [assigned_profile]
-        result.provenance = provenance
-
-        trace = dict(result.retrieval_trace or {}) if isinstance(result.retrieval_trace, dict) else {}
-        trace.setdefault("row_distribution_profile", assigned_profile)
-        trace.setdefault("provider_failover_chain", [assigned_profile])
-        trace.setdefault("provider_failover_final_provider", assigned_profile)
-        result.retrieval_trace = trace
-        return result
-
-    async def _invoke_row_evaluator(
-        self,
-        *,
-        check: AuditCheck,
-        row_id: str,
-        row_index: int,
-        evidence_context: str,
-        evidence_refs: list[str],
-        project_profile: dict[str, Any] | None,
-        provider_health_memory: Any | None = None,
-    ) -> RowEvaluationResult:
-        assigned_profile = self._assigned_repo_profile_for_row(row_index)
-        evaluator = self.evaluator
-        if assigned_profile is not None:
-            evaluator = self._build_row_assigned_evaluator(assigned_profile)
-
-        result = await evaluator.aevaluate_row(
-            check=check,
-            evidence=None,
-            row_id=row_id,
-            row_index=row_index,
-            evidence_context=evidence_context,
-            evidence_refs=evidence_refs,
-            project_profile=project_profile,
-            provider_health_memory=provider_health_memory,
-        )
-        if assigned_profile is not None:
-            result = self._apply_assigned_profile_provenance(result, assigned_profile=assigned_profile)
-        provider_name = str(getattr(getattr(result, "provenance", None), "final_provider", None) or "").strip()
-        if not provider_name:
-            provider_name = str(getattr(getattr(result, "provenance", None), "original_provider", None) or "").strip()
-        if not provider_name:
-            provider_name = assigned_profile
-        await self._record_runtime_provider_result(
-            cast("ProviderHealthMemory | None", provider_health_memory),
-            provider_name=provider_name,
-            result=result,
-        )
-        return result
-
-    # ---------------------------------------------------------------------------
-    # FR-145.2: Post-pass ERROR retry sweep
-    # ---------------------------------------------------------------------------
-
-    def _select_healthy_provider(
-        self,
-        failover_profiles: list[str],
-        health_memory: ProviderHealthMemory,
-        *,
-        exclude: set[str],
-    ) -> str | None:
-        """Select the next healthy provider, skipping cooled-down and excluded ones.
-
-        Args:
-            failover_profiles: Ordered list of candidate profile names.
-            health_memory: Shared health state accumulated during the primary pass.
-            exclude: Provider names to skip (original + already tried).
-
-        Returns:
-            A profile name to retry with, or None if all are blocked/cooled down.
-        """
-        for candidate in failover_profiles:
-            if candidate in exclude:
-                continue
-            if health_memory.should_skip_provider(candidate):
-                continue
-            return candidate
-        return None
-
-    def _build_retry_evaluator(self, profile: str) -> RowEvaluator:
-        """Build a RowEvaluator configured to use the given profile for retry."""
-        return self._build_row_assigned_evaluator(profile)
-
-    @staticmethod
-    def _is_timeout_failoverable(timeout_kind: str) -> bool:
-        """Return True if the timeout kind should trigger failover (TSK-145.13).
-
-        TIMEOUT_WITH_GROUNDING must NEVER trigger failover — it already has
-        partial grounding and a FAIL result. STALL, NO_PROGRESS, and
-        CHURN_DETECTED indicate that the primary provider was unable to make
-        durable forward progress, so a second provider may succeed.
-        """
-        return timeout_kind in {
-            TimeoutKind.TIMEOUT_PROVIDER_STALL.value,
-            TimeoutKind.TIMEOUT_NO_PROGRESS.value,
-            TimeoutKind.TIMEOUT_CHURN_DETECTED.value,
-        }
-
-    async def _attempt_timeout_failover(
-        self,
-        *,
-        timeout_result: RowEvaluationResult,
-        check: AuditCheck,
-        row_index: int,
-        evidence_context: str,
-        evidence_refs: list[str],
-        project_profile: dict[str, Any] | None,
-        health_memory: ProviderHealthMemory,
-        failover_profiles: list[str],
-        stats: _ResilienceRetryStats | None = None,
-    ) -> RowEvaluationResult:
-        """Attempt ONE failover evaluation for a timed-out row (TSK-145.13/14).
-
-        Called BEFORE the timeout result is finalised as ERROR.  Returns the
-        failover result on success, or the original timeout_result unchanged if
-        failover is not possible or also fails.
-
-        Rules:
-        - Only TIMEOUT_PROVIDER_STALL and TIMEOUT_NO_PROGRESS are eligible.
-        - TIMEOUT_WITH_GROUNDING is never retried (has partial grounding).
-        - Exactly one failover attempt is made (no retry loop).
-        - Providers already in the failover chain are excluded.
-        """
-        # Extract the timeout_kind from the retrieval trace.
-        timeout_trace = dict(timeout_result.retrieval_trace or {})
-        timeout_kind = str(timeout_trace.get("timeout_kind") or "").strip()
-
-        if not timeout_kind or not self._is_timeout_failoverable(timeout_kind):
-            return timeout_result
-
-        if not failover_profiles:
-            return timeout_result
-
-        # Determine providers to exclude: original provider + any prior chain.
-        original_provider = str(
-            getattr(getattr(timeout_result, "provenance", None), "original_provider", None) or ""
-        ).strip()
-        chain = list(getattr(getattr(timeout_result, "provenance", None), "failover_chain", None) or [])
-        exclude: set[str] = set(filter(None, [original_provider, *chain]))
-
-        next_provider = self._select_healthy_provider(
-            failover_profiles,
-            health_memory,
-            exclude=exclude,
-        )
-
-        # Phase 146 (FR-146.6): Bounded wait-and-resume when all providers are
-        # temporarily exhausted during timeout failover (TSK-146.10).
-        if next_provider is None:
-            exhaustion_reason = _all_providers_exhausted_reason(
-                health_memory,
-                failover_profiles=failover_profiles,
-            )
-            if exhaustion_reason == "quota_all_providers_exhausted":
-                quota_wait_max = _quota_wait_max_seconds()
-                if quota_wait_max > 0:
-                    recovered = await self._bounded_wait_for_provider_recovery(
-                        health_memory,
-                        failover_profiles=failover_profiles,
-                        max_wait_seconds=quota_wait_max,
-                        row_id=timeout_result.row_id,
-                        check_id=timeout_result.check_id,
-                    )
-                    if recovered is not None and recovered not in exclude:
-                        next_provider = recovered
-
-            if next_provider is None:
-                # Annotate timeout result with exhaustion metadata (AC-146.6.4 / TSK-146.11).
-                self._annotate_quota_exhaustion(
-                    timeout_result,
-                    health_memory=health_memory,
-                    failover_profiles=failover_profiles,
-                    wait_attempted=exhaustion_reason == "quota_all_providers_exhausted"
-                    and _quota_wait_max_seconds() > 0,
-                    wait_expired=True,
-                )
-                logger.debug(
-                    "timeout_failover_no_healthy_provider",
-                    row_id=timeout_result.row_id,
-                    check_id=timeout_result.check_id,
-                    timeout_kind=timeout_kind,
-                    exclude=sorted(exclude),
-                    exhaustion_reason=exhaustion_reason,
-                )
-                return timeout_result
-
-        # Build RowFailoverContext with partial evidence from the timed-out result (TSK-145.14).
-        progress_snapshot = timeout_trace.get("timeout_progress_snapshot")
-        partial_evidence_refs: list[dict[str, Any]] = []
-        if isinstance(progress_snapshot, dict):
-            raw_refs = progress_snapshot.get("evidence_refs")
-            if isinstance(raw_refs, list):
-                partial_evidence_refs = [{"ref": str(r)} for r in raw_refs if r]
-        failover_context = RowFailoverContext(
-            row_id=timeout_result.row_id,
-            check_id=timeout_result.check_id,
-            evidence_refs=partial_evidence_refs,
-            route_state={},
-            partial_progress={
-                "timeout_kind": timeout_kind,
-                "timeout_snapshot": dict(progress_snapshot) if isinstance(progress_snapshot, dict) else {},
-            },
-            original_provider=original_provider or None,
-            failover_count=len(chain),
-            failover_chain=chain + ([original_provider] if original_provider else []),
-            timeout_telemetry_snapshot={
-                k: v
-                for k, v in timeout_trace.items()
-                if k in {"timeout_kind", "timeout_scope", "stall_duration_seconds", "stall_reason"}
-            },
-        )
-
-        if stats is not None and next_provider not in stats.providers_used:
-            stats.providers_used.append(next_provider)
-        if stats is not None:
-            stats.failover_hops += 1
-
-        # Evidence refs for the re-evaluation call.
-        evidence_refs_list: list[str] = list(
-            (timeout_result.retry_metadata or {}).get("evidence_refs", []) or evidence_refs or []
-        )
-        evidence_context_str = str(timeout_trace.get("evidence_context_str", "") or evidence_context)
-
-        logger.info(
-            "timeout_failover_attempt",
-            row_id=timeout_result.row_id,
-            check_id=timeout_result.check_id,
-            timeout_kind=timeout_kind,
-            original_provider=original_provider or "unknown",
-            failover_provider=next_provider,
-            failover_context_partial_refs=len(failover_context.evidence_refs),
-        )
-
-        try:
-            failover_evaluator = self._build_retry_evaluator(next_provider)
-            failover_result = await failover_evaluator.aevaluate_row(
-                check=check,
-                evidence=None,
-                row_id=timeout_result.row_id,
-                row_index=row_index,
-                evidence_context=evidence_context_str,
-                evidence_refs=evidence_refs_list,
-                project_profile=project_profile,
-            )
-            failover_result = self._apply_assigned_profile_provenance(
-                failover_result,
-                assigned_profile=next_provider,
-            )
-        except Exception as exc:
-            logger.warning(
-                "timeout_failover_failed",
-                row_id=timeout_result.row_id,
-                check_id=timeout_result.check_id,
-                failover_provider=next_provider,
-                error=str(exc),
-            )
-            return timeout_result
-
-        if failover_result.status == RowStatus.ERROR:
-            logger.debug(
-                "timeout_failover_still_error",
-                row_id=timeout_result.row_id,
-                check_id=timeout_result.check_id,
-                failover_provider=next_provider,
-            )
-            return timeout_result
-
-        # Failover succeeded — annotate the result with failover provenance.
-        failover_result.retry_count = timeout_result.retry_count + 1
-        failover_result.retry_metadata = {
-            **(failover_result.retry_metadata or {}),
-            "timeout_failover_pass": True,
-            "timeout_failover_original_provider": original_provider or "unknown",
-            "timeout_failover_provider": next_provider,
-            "timeout_failover_kind": timeout_kind,
-            "timeout_failover_context": {
-                "failover_count": failover_context.failover_count,
-                "failover_chain": failover_context.failover_chain,
-                "partial_progress": failover_context.partial_progress,
-            },
-        }
-        if stats is not None:
-            stats.retry_successes += 1
-
-        logger.info(
-            "timeout_failover_succeeded",
-            row_id=timeout_result.row_id,
-            check_id=timeout_result.check_id,
-            timeout_kind=timeout_kind,
-            original_provider=original_provider or "unknown",
-            failover_provider=next_provider,
-            new_status=failover_result.status.value,
-        )
-        return failover_result
-
-    async def _retry_error_rows(
-        self,
-        results: list[RowEvaluationResult],
-        health_memory: ProviderHealthMemory,
-        budget: ResilienceBudget,
-        failover_profiles: list[str],
-        project_profile: dict[str, Any] | None,
-        stats: _ResilienceRetryStats | None = None,
-    ) -> list[RowEvaluationResult]:
-        """Post-pass retry sweep for failoverable ERROR rows (FR-145.2).
-
-        Iterates results from the primary pass. For each ERROR row that is
-        failoverable and within budget, re-evaluates using a healthy alternative
-        provider. Replaces the original result only on success.
-
-        The ``retry_count`` on the replacement result is incremented to signal
-        that a retry occurred. The ``retry_metadata`` carries provenance of the
-        retry (``resilience_retry_pass``, original provider, new provider).
-
-        Args:
-            results: Primary-pass results (not mutated; returns a new list).
-            health_memory: Shared ProviderHealthMemory from the primary pass.
-            budget: ResilienceBudget instance (mutated in place via consume()).
-            failover_profiles: Ordered list of alternative provider names.
-            project_profile: Project-level profile dict passed to the evaluator.
-            stats: Optional accumulator for resilience telemetry (FR-145.7).
-
-        Returns:
-            New list with ERROR rows replaced where retry succeeded.
-        """
-        if not failover_profiles or budget.remaining <= 0:
-            return list(results)
-
-        updated = list(results)
-
-        # Seed stats with pre-retry snapshot so callers get before/after counts.
-        if stats is not None:
-            stats.error_rows_before_retry = sum(1 for r in results if r.status == RowStatus.ERROR)
-            stats.budget_total = budget.total
-
-        for i, result in enumerate(results):
-            if result.status != RowStatus.ERROR:
-                continue
-
-            # Determine if this error is failoverable via reason_code in retry_metadata.
-            classification_reason = str((result.retry_metadata or {}).get("reason_code", "")).strip().lower()
-            if classification_reason in _NON_FAILOVERABLE_REASONS:
-                logger.debug(
-                    "resilience_retry_skipped_non_failoverable",
-                    row_id=result.row_id,
-                    check_id=result.check_id,
-                    reason_code=classification_reason,
-                )
-                continue
-
-            # Check budget before attempting retry.
-            if not budget.consume():
-                logger.info(
-                    "resilience_budget_exhausted",
-                    remaining_errors=sum(1 for r in results[i:] if r.status == RowStatus.ERROR),
-                    calls_used=budget.calls_used,
-                    budget_total=budget.total,
-                )
-                break
-
-            if stats is not None:
-                stats.retry_attempts += 1
-                stats.budget_used = budget.calls_used
-
-            # Determine which providers to exclude: original + any already tried.
-            original_provider = str(getattr(result.provenance, "original_provider", None) or "").strip()
-            chain = list(getattr(result.provenance, "failover_chain", None) or [])
-            exclude: set[str] = set(filter(None, [original_provider, *chain]))
-
-            next_provider = self._select_healthy_provider(
-                failover_profiles,
-                health_memory,
-                exclude=exclude,
-            )
-
-            # Phase 146 (FR-146.6): Bounded wait-and-resume when all providers are
-            # temporarily exhausted but not permanently blocked (TSK-146.10).
-            wait_attempted = False
-            wait_expired = False
-            if next_provider is None:
-                exhaustion_reason = _all_providers_exhausted_reason(
-                    health_memory,
-                    failover_profiles=failover_profiles,
-                )
-                if exhaustion_reason == "quota_all_providers_exhausted":
-                    # All providers are in temporary cooldown — attempt bounded wait.
-                    quota_wait_max = _quota_wait_max_seconds()
-                    if quota_wait_max > 0:
-                        wait_attempted = True
-                        recovered = await self._bounded_wait_for_provider_recovery(
-                            health_memory,
-                            failover_profiles=failover_profiles,
-                            max_wait_seconds=quota_wait_max,
-                            row_id=result.row_id,
-                            check_id=result.check_id,
-                        )
-                        if recovered is not None and recovered not in exclude:
-                            next_provider = recovered
-                        else:
-                            wait_expired = True
-                    else:
-                        wait_expired = True
-
-                if next_provider is None:
-                    # Annotate with exhaustion metadata (AC-146.6.4 / TSK-146.11).
-                    self._annotate_quota_exhaustion(
-                        result,
-                        health_memory=health_memory,
-                        failover_profiles=failover_profiles,
-                        wait_attempted=wait_attempted,
-                        wait_expired=wait_expired,
-                    )
-                    logger.debug(
-                        "resilience_retry_no_healthy_provider",
-                        row_id=result.row_id,
-                        check_id=result.check_id,
-                        exclude=sorted(exclude),
-                        exhaustion_reason=exhaustion_reason,
-                        wait_attempted=wait_attempted,
-                        wait_expired=wait_expired,
-                    )
-                    continue
-
-            # Build RowFailoverContext carrying forward evidence from the original result.
-            failover_context = RowFailoverContext(
-                row_id=result.row_id,
-                check_id=result.check_id,
-                evidence_refs=list(
-                    (result.retry_metadata or {}).get("evidence_refs", [])
-                    or (getattr(result, "evidence_anchors", None) and [str(a) for a in (result.evidence_anchors or [])])
-                    or []
-                ),
-                route_state=dict((result.retry_metadata or {}).get("route_state", {})),
-                partial_progress=dict((result.retry_metadata or {}).get("partial_progress", {})),
-                original_provider=original_provider or None,
-                failover_count=len(chain),
-                failover_chain=chain + ([original_provider] if original_provider else []),
-                timeout_telemetry_snapshot={},
-            )
-
-            retry_evaluator = self._build_retry_evaluator(next_provider)
-            if stats is not None and next_provider not in stats.providers_used:
-                stats.providers_used.append(next_provider)
-            if stats is not None:
-                stats.failover_hops += 1
-
-            # Reconstruct evidence for the row from the original retrieval trace.
-            evidence_context_str = str((result.retrieval_trace or {}).get("evidence_context_str", ""))
-            evidence_refs_list: list[str] = list((result.retry_metadata or {}).get("evidence_refs", []) or [])
-
-            # Determine row_index: extract from row_id "check_id:row_N" pattern.
-            try:
-                row_index = int(result.row_id.split(":row_")[-1])
-            except (ValueError, IndexError):
-                row_index = 0
-
-            # Find the original AuditCheck for this result.
-            check_match: AuditCheck | None = None
-            for _section_id, check in self._checks:
-                if check.id == result.check_id:
-                    check_match = check
-                    break
-            if check_match is None:
-                logger.debug(
-                    "resilience_retry_check_not_found",
-                    row_id=result.row_id,
-                    check_id=result.check_id,
-                )
-                continue
-
-            logger.info(
-                "resilience_retry_attempt",
-                row_id=result.row_id,
-                check_id=result.check_id,
-                original_provider=original_provider or "unknown",
-                retry_provider=next_provider,
-                budget_calls_used=budget.calls_used,
-                budget_total=budget.total,
-            )
-
-            try:
-                retry_result = await retry_evaluator.aevaluate_row(
-                    check=check_match,
-                    evidence=None,
-                    row_id=result.row_id,
-                    row_index=row_index,
-                    evidence_context=evidence_context_str,
-                    evidence_refs=evidence_refs_list,
-                    project_profile=project_profile,
-                )
-                retry_result = self._apply_assigned_profile_provenance(
-                    retry_result,
-                    assigned_profile=next_provider,
-                )
-            except Exception as exc:
-                logger.warning(
-                    "resilience_retry_failed",
-                    row_id=result.row_id,
-                    check_id=result.check_id,
-                    retry_provider=next_provider,
-                    error=str(exc),
-                )
-                continue
-
-            if retry_result.status == RowStatus.ERROR:
-                logger.debug(
-                    "resilience_retry_still_error",
-                    row_id=result.row_id,
-                    check_id=result.check_id,
-                    retry_provider=next_provider,
-                )
-                continue
-
-            # Retry succeeded — replace the original result and annotate.
-            retry_result.retry_count = result.retry_count + 1
-            retry_result.retry_metadata = {
-                **(retry_result.retry_metadata or {}),
-                "resilience_retry_pass": True,
-                "resilience_original_provider": original_provider or "unknown",
-                "resilience_retry_provider": next_provider,
-                "resilience_failover_context": {
-                    "failover_count": failover_context.failover_count,
-                    "failover_chain": failover_context.failover_chain,
-                },
-            }
-            updated[i] = retry_result
-            if stats is not None:
-                stats.retry_successes += 1
-            logger.info(
-                "resilience_retry_succeeded",
-                row_id=result.row_id,
-                check_id=result.check_id,
-                original_provider=original_provider or "unknown",
-                retry_provider=next_provider,
-                new_status=retry_result.status.value,
-            )
-
-        # Sync final budget_used after the loop (some budget slots may be consumed
-        # without a matching attempt when budget is already 0 on entry).
-        if stats is not None:
-            stats.budget_used = budget.calls_used
-            # Phase 146 telemetry: count rows with quota-exhaustion annotation.
-            stats.quota_exhausted_rows = sum(
-                1
-                for r in updated
-                if r.status == RowStatus.ERROR
-                and isinstance(r.retry_metadata, dict)
-                and r.retry_metadata.get("quota_all_providers_exhausted")
-            )
-            stats.quota_wait_attempted_rows = sum(
-                1
-                for r in updated
-                if isinstance(r.retry_metadata, dict) and r.retry_metadata.get("quota_exhaustion_wait_attempted")
-            )
-            stats.quota_wait_recovered_rows = sum(
-                1
-                for r in updated
-                if isinstance(r.retry_metadata, dict) and r.retry_metadata.get("quota_exhaustion_wait_recovered")
-            )
-
-        return updated
-
-    # ---------------------------------------------------------------------------
-    # TSK-145.9/145.10/145.11: Post-pass low-confidence re-evaluation sweep
-    # ---------------------------------------------------------------------------
-
-    async def _reevaluate_low_confidence_rows(
-        self,
-        results: list[RowEvaluationResult],
-        health_memory: ProviderHealthMemory,
-        budget: ResilienceBudget,
-        failover_profiles: list[str],
-        project_profile: dict[str, Any] | None,
-    ) -> list[RowEvaluationResult]:
-        """Post-pass re-evaluation sweep for PARTIAL or low-score rows (TSK-145.9).
-
-        Runs AFTER ``_retry_error_rows``, sharing the same ``ResilienceBudget``
-        instance.  If the ERROR retry sweep exhausted the budget, no re-evals
-        happen (budget.remaining == 0).
-
-        Eligible rows:
-        - status == PARTIAL, OR
-        - score_1_5 is not None and score_1_5 <= threshold
-          (default threshold: 2; override via
-           ``VDS_AUDIT_LOW_CONFIDENCE_REEVAL_THRESHOLD`` env var).
-
-        For each eligible row the method:
-        1. Consumes one budget slot (skips if exhausted).
-        2. Selects a healthy provider, preferring one different from the
-           original.
-        3. Re-evaluates using the same ``aevaluate_row`` interface but with a
-           ``prior_result_context`` appended to the evidence so the model is
-           instructed to critically re-examine rather than anchor on the prior
-           assessment (TSK-145.10).
-        4. Replaces the result ONLY if ``reeval_score_1_5 > original_score_1_5``
-           (TSK-145.11); otherwise keeps the original.
-
-        Args:
-            results: Primary-pass results (after ERROR retry sweep); not
-                mutated — returns a new list.
-            health_memory: Shared ProviderHealthMemory.
-            budget: ResilienceBudget instance shared with ERROR retry sweep;
-                mutated in place via consume().
-            failover_profiles: Ordered list of alternative provider names.
-            project_profile: Project-level profile dict passed to evaluator.
-
-        Returns:
-            New list with low-confidence rows replaced where re-eval improved
-            the score.
-        """
-        import os as _os
-
-        if budget.remaining <= 0:
-            return list(results)
-
-        # Parse threshold from env (default 2 → score_1_5 in {1, 2} are eligible).
-        try:
-            threshold = int((_os.getenv("VDS_AUDIT_LOW_CONFIDENCE_REEVAL_THRESHOLD") or "2").strip())
-        except ValueError:
-            threshold = 2
-
-        updated = list(results)
-
-        for i, result in enumerate(results):
-            # Eligibility: PARTIAL status OR low score_1_5.
-            is_partial = result.status == RowStatus.PARTIAL
-            score_1_5 = result.score_1_5
-            is_low_score = score_1_5 is not None and score_1_5 <= threshold
-
-            if not (is_partial or is_low_score):
-                continue
-
-            # Check budget before attempting re-eval.
-            if not budget.consume():
-                logger.info(
-                    "low_confidence_reeval_budget_exhausted",
-                    remaining_eligible=sum(
-                        1
-                        for r in results[i:]
-                        if r.status == RowStatus.PARTIAL or (r.score_1_5 is not None and r.score_1_5 <= threshold)
-                    ),
-                    calls_used=budget.calls_used,
-                    budget_total=budget.total,
-                )
-                break
-
-            # Prefer a provider different from the original.
-            original_provider = str(getattr(result.provenance, "original_provider", None) or "").strip()
-            chain = list(getattr(result.provenance, "failover_chain", None) or [])
-            exclude: set[str] = set(filter(None, [original_provider, *chain]))
-
-            next_provider = self._select_healthy_provider(
-                failover_profiles,
-                health_memory,
-                exclude=exclude,
-            )
-            # Fall back to any provider (including original) if all are excluded.
-            if next_provider is None and failover_profiles:
-                next_provider = self._select_healthy_provider(
-                    failover_profiles,
-                    health_memory,
-                    exclude=set(),
-                )
-            if next_provider is None:
-                logger.debug(
-                    "low_confidence_reeval_no_healthy_provider",
-                    row_id=result.row_id,
-                    check_id=result.check_id,
-                )
-                continue
-
-            # Find the original AuditCheck for this result.
-            check_match = None
-            for _section_id, check in self._checks:
-                if check.id == result.check_id:
-                    check_match = check
-                    break
-            if check_match is None:
-                logger.debug(
-                    "low_confidence_reeval_check_not_found",
-                    row_id=result.row_id,
-                    check_id=result.check_id,
-                )
-                continue
-
-            # Determine row_index from row_id "check_id:row_N" pattern.
-            try:
-                row_index = int(result.row_id.split(":row_")[-1])
-            except (ValueError, IndexError):
-                row_index = 0
-
-            # Reconstruct evidence context from the original retrieval trace.
-            evidence_context_str = str((result.retrieval_trace or {}).get("evidence_context_str", ""))
-            evidence_refs_list: list[str] = list((result.retry_metadata or {}).get("evidence_refs", []) or [])
-
-            # Build prior_result_context to instruct the model not to anchor
-            # on the prior assessment (TSK-145.10).
-            prior_result_context: dict[str, Any] = {
-                "prior_score_1_5": score_1_5,
-                "prior_status": result.status.value,
-                "prior_reason": result.reason or "",
-                "prior_evidence_refs": evidence_refs_list[:20],
-                "reeval_instruction": (
-                    "Critically re-examine the evidence. "
-                    "Do not simply agree with the prior assessment. "
-                    "The previous evaluation may have been incomplete or anchored incorrectly."
-                ),
-            }
-
-            # Append prior context to evidence_context so the evaluator prompt
-            # carries re-eval framing without requiring interface changes.
-            reeval_evidence_context = (
-                f"{evidence_context_str}\n\n"
-                "[Re-evaluation context]\n"
-                f"Prior score (1-5): {prior_result_context['prior_score_1_5']}\n"
-                f"Prior status: {prior_result_context['prior_status']}\n"
-                f"Prior reason: {prior_result_context['prior_reason']}\n"
-                f"{prior_result_context['reeval_instruction']}"
-            )
-
-            reeval_evaluator = self._build_retry_evaluator(next_provider)
-
-            logger.info(
-                "low_confidence_reeval_attempted",
-                row_id=result.row_id,
-                check_id=result.check_id,
-                original_provider=original_provider or "unknown",
-                reeval_provider=next_provider,
-                prior_score_1_5=score_1_5,
-                prior_status=result.status.value,
-                budget_calls_used=budget.calls_used,
-                budget_total=budget.total,
-            )
-
-            try:
-                reeval_result = await reeval_evaluator.aevaluate_row(
-                    check=check_match,
-                    evidence=None,
-                    row_id=result.row_id,
-                    row_index=row_index,
-                    evidence_context=reeval_evidence_context,
-                    evidence_refs=evidence_refs_list,
-                    project_profile=project_profile,
-                )
-                reeval_result = self._apply_assigned_profile_provenance(
-                    reeval_result,
-                    assigned_profile=next_provider,
-                )
-            except Exception as exc:
-                logger.warning(
-                    "low_confidence_reeval_failed",
-                    row_id=result.row_id,
-                    check_id=result.check_id,
-                    reeval_provider=next_provider,
-                    error=str(exc),
-                )
-                continue
-
-            # TSK-145.11: Replace only if re-eval score is strictly higher.
-            original_score = result.score_1_5 or 0
-            reeval_score = reeval_result.score_1_5 or 0
-
-            if reeval_score > original_score:
-                reeval_result.retry_count = result.retry_count + 1
-                reeval_result.retry_metadata = {
-                    **(reeval_result.retry_metadata or {}),
-                    "low_confidence_reeval_pass": True,
-                    "low_confidence_original_provider": original_provider or "unknown",
-                    "low_confidence_reeval_provider": next_provider,
-                    "low_confidence_prior_score_1_5": original_score,
-                    "low_confidence_prior_result_context": prior_result_context,
-                }
-                updated[i] = reeval_result
-                logger.info(
-                    "low_confidence_reeval_improved",
-                    row_id=result.row_id,
-                    check_id=result.check_id,
-                    original_score=original_score,
-                    new_score=reeval_score,
-                    reeval_provider=next_provider,
-                )
-            else:
-                logger.info(
-                    "low_confidence_reeval_kept_original",
-                    row_id=result.row_id,
-                    check_id=result.check_id,
-                    original_score=original_score,
-                    reeval_score=reeval_score,
-                )
-
-        return updated
-
-    def _build_timeout_result_with_trace(
-        self,
-        *,
-        row_id: str,
-        check_id: str,
-        message: str,
-        progress: dict[str, Any] | None,
-        timeout_trace: dict[str, Any],
-        matcher_retrieval_trace: dict[str, Any] | None,
-    ) -> RowEvaluationResult:
-        progress = self._sanitize_timeout_progress_snapshot_for_app_config_only(
-            progress,
-            matcher_retrieval_trace=matcher_retrieval_trace,
-        )
-        timeout_trace = self._sanitize_timeout_trace_for_app_config_only(
-            timeout_trace,
-            progress=progress,
-            matcher_retrieval_trace=matcher_retrieval_trace,
-        )
-        degraded_timeout_finalize = bool(timeout_trace.get("timeout_degraded_finalize"))
-        result = self._create_timeout_result(
-            row_id,
-            check_id,
-            message,
-            degraded_timeout_finalize=degraded_timeout_finalize,
-            progress=progress,
-        )
-        self._attach_timeout_progress_telemetry(
-            timeout_trace=timeout_trace,
-            progress=progress,
-            row_id=row_id,
-            check_id=check_id,
-        )
-        base_trace = _sanitize_retrieval_trace(matcher_retrieval_trace) or {}
-        result.retrieval_trace = _sanitize_retrieval_trace({**base_trace, **timeout_trace})
-        result = RowEvaluator._sanitize_app_config_only_result(result)
-        return RowEvaluator._salvage_app_config_only_timeout_result(result)
-
-    @staticmethod
-    def _build_timeout_progress_config_rows(
-        *,
-        progress: dict[str, Any] | None,
-        matcher_retrieval_trace: dict[str, Any] | None,
-    ) -> list[dict[str, Any]]:
-        config_rows: list[dict[str, Any]] = []
-        for container in (
-            matcher_retrieval_trace,
-            (matcher_retrieval_trace or {}).get("tool_first_loop")
-            if isinstance(matcher_retrieval_trace, dict)
-            else None,
-            progress,
-        ):
-            if not isinstance(container, dict):
-                continue
-            raw_rows = container.get("config_repo_companion_refs")
-            if not isinstance(raw_rows, list):
-                continue
-            for row in raw_rows:
-                if isinstance(row, dict):
-                    config_rows.append(dict(row))
-        if config_rows:
-            return config_rows
-
-        fallback_rows: list[dict[str, Any]] = []
-        raw_refs = progress.get("evidence_refs") if isinstance(progress, dict) else None
-        for ref in raw_refs if isinstance(raw_refs, list) else []:
-            ref_value = RowEvaluator._normalize_trace_ref(ref)
-            if not RowEvaluator._is_config_ref_value(ref_value):
-                continue
-            parts = Path(ref_value).parts
-            env_name = str(parts[0]).strip().lower() if len(parts) >= 2 else ""
-            fallback_rows.append(
-                {
-                    "ref_value": ref_value,
-                    "ref": ref_value,
-                    "source_repo": "app-config",
-                    "metadata": {
-                        "env": env_name,
-                        "source_repo_slug": "app-config",
-                        "config_root": "app-config",
-                    },
-                }
-            )
-        return fallback_rows
-
-    @classmethod
-    def _compute_timeout_app_config_sanitization(
-        cls,
-        *,
-        progress: dict[str, Any] | None,
-        matcher_retrieval_trace: dict[str, Any] | None,
-    ) -> tuple[dict[str, Any] | None, list[dict[str, Any]], set[str]]:
-        if not isinstance(progress, dict):
-            return None, [], set()
-        interpretation = (
-            progress.get("requirement_interpretation")
-            if isinstance(progress.get("requirement_interpretation"), dict)
-            else {}
-        )
-        sanitized_interpretation = RowEvaluator._sanitize_requirement_interpretation_for_app_config_only(interpretation)
-        if not RowEvaluator._is_app_config_only_requirement(
-            requirement_interpretation=sanitized_interpretation,
-        ):
-            return None, [], set()
-
-        raw_refs = progress.get("evidence_refs") if isinstance(progress.get("evidence_refs"), list) else []
-        requirement_text = str(
-            sanitized_interpretation.get("intent") or progress.get("requirement_intent") or ""
-        ).strip()
-        requirement_guidance = str(sanitized_interpretation.get("detailed_guidance") or "").strip() or None
-        config_rows = cls._build_timeout_progress_config_rows(
-            progress=progress,
-            matcher_retrieval_trace=matcher_retrieval_trace,
-        )
-        selected_rows = RowEvaluator._select_app_config_only_rows(
-            requirement_text=requirement_text,
-            requirement_guidance=requirement_guidance,
-            requirement_interpretation=sanitized_interpretation,
-            config_rows=config_rows,
-        )
-        allowed_ref_values = RowEvaluator._allowed_app_config_only_ref_values(
-            requirement_text=requirement_text,
-            requirement_guidance=requirement_guidance,
-            requirement_interpretation=sanitized_interpretation,
-            config_rows=config_rows,
-            fallback_refs=raw_refs,
-        )
-        return sanitized_interpretation, selected_rows, allowed_ref_values
-
-    @classmethod
-    def _sanitize_timeout_progress_snapshot_for_app_config_only(
-        cls,
-        progress: dict[str, Any] | None,
-        *,
-        matcher_retrieval_trace: dict[str, Any] | None,
-    ) -> dict[str, Any] | None:
-        if not isinstance(progress, dict):
-            return progress
-        sanitized_interpretation, selected_rows, allowed_ref_values = cls._compute_timeout_app_config_sanitization(
-            progress=progress,
-            matcher_retrieval_trace=matcher_retrieval_trace,
-        )
-        if sanitized_interpretation is None:
-            return progress
-
-        normalized = dict(progress)
-        normalized["requirement_interpretation"] = sanitized_interpretation
-        normalized["evidence_refs"] = RowEvaluator._filter_refs_for_app_config_only(
-            normalized.get("evidence_refs"),
-            allowed_ref_values=allowed_ref_values,
-        )
-        normalized["config_repo_companion_refs"] = selected_rows
-        return normalized
-
-    @classmethod
-    def _sanitize_timeout_trace_for_app_config_only(
-        cls,
-        timeout_trace: dict[str, Any],
-        *,
-        progress: dict[str, Any] | None,
-        matcher_retrieval_trace: dict[str, Any] | None,
-    ) -> dict[str, Any]:
-        normalized = dict(timeout_trace)
-        sanitized_progress = cls._sanitize_timeout_progress_snapshot_for_app_config_only(
-            progress,
-            matcher_retrieval_trace=matcher_retrieval_trace,
-        )
-        normalized["timeout_progress_snapshot"] = sanitized_progress
-        sanitized_interpretation, selected_rows, allowed_ref_values = cls._compute_timeout_app_config_sanitization(
-            progress=sanitized_progress,
-            matcher_retrieval_trace=matcher_retrieval_trace,
-        )
-        if sanitized_interpretation is None:
-            return normalized
-
-        filtered_refs = (
-            RowEvaluator._filter_refs_for_app_config_only(
-                sanitized_progress.get("evidence_refs") if isinstance(sanitized_progress, dict) else [],
-                allowed_ref_values=allowed_ref_values,
-            )
-            if isinstance(sanitized_progress, dict)
-            else []
-        )
-        preserved_trace_steps: list[dict[str, Any]] = []
-        progress_loop_trace = (
-            sanitized_progress.get("tool_first_loop")
-            if isinstance(sanitized_progress, dict) and isinstance(sanitized_progress.get("tool_first_loop"), dict)
-            else {}
-        )
-        raw_trace_steps = progress_loop_trace.get("trace_steps") if isinstance(progress_loop_trace, dict) else None
-        if isinstance(raw_trace_steps, list):
-            for step in raw_trace_steps:
-                if not isinstance(step, dict):
-                    continue
-                tool_name = str(step.get("tool") or "").strip()
-                output = step.get("output")
-                updated_step = dict(step)
-                if tool_name == "read_code_file" and isinstance(output, dict):
-                    path_value = RowEvaluator._normalize_trace_ref(
-                        output.get("path") or output.get("file_path") or output.get("source_path")
-                    )
-                    if not RowEvaluator._is_config_ref_value(path_value):
-                        continue
-                    if allowed_ref_values and path_value not in allowed_ref_values:
-                        continue
-                elif tool_name == "record_evidence_refs" and isinstance(output, dict):
-                    sanitized_output = dict(output)
-                    for key in ("refs", "evidence_refs"):
-                        if isinstance(output.get(key), list):
-                            sanitized_output[key] = RowEvaluator._filter_trace_entries_for_app_config_only(
-                                output.get(key),
-                                allowed_ref_values=allowed_ref_values,
-                            )
-                    updated_step["output"] = sanitized_output
-                preserved_trace_steps.append(updated_step)
-        normalized["requirement_interpretation"] = sanitized_interpretation
-        normalized["requirement_intent"] = str(
-            sanitized_interpretation.get("intent") or normalized.get("requirement_intent") or ""
-        ).strip()
-        normalized["evidence_refs"] = filtered_refs
-        normalized["config_repo_companion_refs"] = selected_rows
-        normalized["cross_repo_evidence_refs"] = []
-        normalized["prior_knowledge_refs"] = []
-        normalized.pop("docs", None)
-        normalized["code"] = {"top_k": [], "hits": []}
-        normalized["tool_first_loop"] = {
-            "evidence_refs": filtered_refs,
-            "config_repo_companion_refs": selected_rows,
-            "cross_repo_evidence_refs": [],
-            "prior_knowledge_refs": [],
-            "trace_steps": preserved_trace_steps,
-        }
-        normalized["evidence_ref_sets"] = {
-            "final_cited_refs": filtered_refs[:50],
-            "final_cited_ref_count": len(filtered_refs),
-            "exploratory_refs": [],
-            "exploratory_ref_count": 0,
-            "dropped_refs": [],
-            "dropped_exploratory_ref_count": 0,
-            "final_cited_ref_modality": {"docs": False, "code": False},
-        }
-        normalized["app_config_only_enforced"] = {
-            "applied": True,
-            "allowed_ref_values": sorted(allowed_ref_values)[:20],
-            "final_anchor_count": 0,
-        }
-        return normalized
-
-    async def _evaluate_row_with_legacy_timeout(
-        self,
-        *,
-        check: AuditCheck,
-        row_id: str,
-        row_index: int,
-        evidence_context: str,
-        evidence_refs: list[str],
-        project_profile: dict[str, Any] | None,
-        thread_id: str,
-        matcher_retrieval_trace: dict[str, Any] | None,
-        provider_health_memory: Any | None = None,
-    ) -> RowEvaluationResult:
-        row_timeout_ms = int(self.config.row_timeout_ms)
-        timeout_retry_attempt = 0
-        last_progress_summary: dict[str, int] | None = None
-        while True:
-            try:
-                async with asyncio.timeout(row_timeout_ms / 1000):
-                    return await self._invoke_row_evaluator(
-                        check=check,
-                        row_id=row_id,
-                        row_index=row_index,
-                        evidence_context=evidence_context,
-                        evidence_refs=evidence_refs,
-                        project_profile=project_profile,
-                        provider_health_memory=provider_health_memory,
-                    )
-            except TimeoutError:
-                raw_progress = self._get_row_runtime_progress(row_id)
-                progress = self._normalize_timeout_progress_snapshot(raw_progress)
-                extension_decision = self._decide_row_timeout_extension(
-                    timeout_retry_attempt=timeout_retry_attempt,
-                    current_timeout_ms=row_timeout_ms,
-                    progress=raw_progress,
-                    previous_progress_summary=last_progress_summary,
-                )
-                if bool(extension_decision.get("allow")):
-                    if isinstance(extension_decision.get("progress_summary"), dict):
-                        last_progress_summary = dict(cast("dict[str, int]", extension_decision.get("progress_summary")))
-                    next_timeout_ms = self._next_row_timeout_ms(row_timeout_ms, timeout_retry_attempt)
-                    timeout_retry_attempt += 1
-                    logger.warning(
-                        "row_timeout_extended_for_effective_progress",
-                        row_id=row_id,
-                        check_id=check.id,
-                        thread_id=thread_id,
-                        run_id=thread_id,
-                        audit_id=thread_id,
-                        timeout_retry_attempt=timeout_retry_attempt,
-                        previous_timeout_ms=row_timeout_ms,
-                        next_timeout_ms=next_timeout_ms,
-                        progress=progress,
-                    )
-                    row_timeout_ms = next_timeout_ms
-                    continue
-                degraded_timeout_finalize = self._should_degrade_timeout_finalize(progress)
-                timeout_trace = {
-                    "timeout_retry_attempts": timeout_retry_attempt,
-                    "timeout_progress_snapshot": progress,
-                    "timeout_extended": timeout_retry_attempt > 0,
-                    "timeout_extension_denial_reason": str(extension_decision.get("reason") or "unknown"),
-                    "timeout_degraded_finalize": degraded_timeout_finalize,
-                    "timeout_kind": self._derive_timeout_kind(
-                        progress=progress,
-                        degraded_timeout_finalize=degraded_timeout_finalize,
-                    ),
-                    "timeout_terminal_status": (RowStatus.FAIL if degraded_timeout_finalize else RowStatus.ERROR).value,
-                }
-                extension_progress_delta = extension_decision.get("progress_delta")
-                if isinstance(extension_progress_delta, dict):
-                    timeout_trace["timeout_extension_progress_delta"] = dict(extension_progress_delta)
-                return self._build_timeout_result_with_trace(
-                    row_id=row_id,
-                    check_id=check.id,
-                    message="Row timeout exceeded",
-                    progress=progress,
-                    timeout_trace=timeout_trace,
-                    matcher_retrieval_trace=matcher_retrieval_trace,
-                )
-
-    def _detect_lease_progress_signal(
-        self,
-        previous_progress: dict[str, Any] | None,
-        current_progress: dict[str, Any] | None,
-    ) -> str | None:
-        if not isinstance(current_progress, dict):
-            return None
-        previous = previous_progress if isinstance(previous_progress, dict) else {}
-
-        previous_seen = self._coerce_progress_timestamp(previous.get("event_last_seen_at"))
-        current_seen = self._coerce_progress_timestamp(current_progress.get("event_last_seen_at"))
-        if current_seen > previous_seen:
-            return "event_stream"
-
-        for key, signal_type in (
-            ("event_tool_calls_started", "tool_call_started"),
-            ("event_tool_calls_completed", "tool_call_completed"),
-            ("event_skill_execution_effective_tool_calls", "skill_effective"),
-            ("event_turn_index", "partial_output"),
-            ("steps_executed", "route_transition"),
-            ("steps_unique_tools", "route_transition"),
-            ("evidence_refs_count", "evidence_growth"),
-        ):
-            if self._coerce_progress_counter(current_progress.get(key)) > self._coerce_progress_counter(
-                previous.get(key)
-            ):
-                return signal_type
-
-        current_summary = self._extract_progress_summary(current_progress)
-        previous_summary = self._extract_progress_summary(previous)
-        progress_delta = self._compute_progress_delta(previous_summary, current_summary)
-        if isinstance(progress_delta, dict) and max(progress_delta.values(), default=0) > 0:
-            return "tool_activity"
-        return None
-
-    def _derive_lease_stall_reason(
-        self,
-        *,
-        elapsed_since_progress_seconds: float,
-        lease_window_seconds: float,
-        stall_detection_seconds: float,
-        progress: dict[str, Any] | None,
-    ) -> str:
-        if elapsed_since_progress_seconds >= lease_window_seconds:
-            return "lease_expired"
-        termination_reason = str((progress or {}).get("termination_reason") or "").strip().lower()
-        if termination_reason in {"idle_after_tool", "post_tool_idle", "idle_post_tool"}:
-            return "idle_after_tool"
-        if termination_reason in {"provider_stall", "stream_stall"}:
-            return "provider_stall"
-        if termination_reason in {"tool_churn", "agent_tool_churn", "churn"}:
-            return "tool_churn"
-        if elapsed_since_progress_seconds >= stall_detection_seconds:
-            return "stall_detection_threshold"
-        return "no_progress"
-
-    async def _evaluate_row_with_lease_timeout(
-        self,
-        *,
-        check: AuditCheck,
-        row_id: str,
-        row_index: int,
-        evidence_context: str,
-        evidence_refs: list[str],
-        project_profile: dict[str, Any] | None,
-        thread_id: str,
-        matcher_retrieval_trace: dict[str, Any] | None,
-        provider_health_memory: Any | None = None,
-    ) -> RowEvaluationResult:
-        import time
-
-        absolute_timeout_seconds = max(0.05, float(self.config.row_absolute_timeout_ms) / 1000.0)
-        lease_window_seconds = max(0.05, float(self.config.row_progress_lease_seconds))
-        stall_detection_seconds = max(
-            0.05,
-            min(float(self.config.row_stall_detection_seconds), lease_window_seconds),
-        )
-        poll_interval_seconds = min(1.0, stall_detection_seconds, lease_window_seconds)
-
-        task = asyncio.create_task(
-            self._invoke_row_evaluator(
-                check=check,
-                row_id=row_id,
-                row_index=row_index,
-                evidence_context=evidence_context,
-                evidence_refs=evidence_refs,
-                project_profile=project_profile,
-                provider_health_memory=provider_health_memory,
-            )
-        )
-        started_at = time.monotonic()
-        last_progress_at = started_at
-        previous_progress: dict[str, Any] | None = None
-        previous_progress_summary: dict[str, int] | None = None
-        liveness_lease_renewals = 0
-        last_progress_signal: float | None = None
-        last_progress_signal_type: str | None = None
-
-        try:
-            while True:
-                now = time.monotonic()
-                remaining_absolute_seconds = absolute_timeout_seconds - (now - started_at)
-                if remaining_absolute_seconds <= 0:
-                    progress = self._normalize_timeout_progress_snapshot(self._get_row_runtime_progress(row_id))
-                    degraded_timeout_finalize = self._should_degrade_timeout_finalize(progress)
-                    stall_duration_seconds = max(0.0, now - last_progress_at)
-                    timeout_trace = {
-                        "timeout_retry_attempts": 0,
-                        "timeout_progress_snapshot": progress,
-                        "timeout_extended": liveness_lease_renewals > 0,
-                        "timeout_extension_denial_reason": "absolute_timeout_reached",
-                        "timeout_degraded_finalize": degraded_timeout_finalize,
-                        "timeout_kind": self._derive_timeout_kind(
-                            progress=progress,
-                            degraded_timeout_finalize=degraded_timeout_finalize,
-                            absolute_timeout_exceeded=True,
-                        ),
-                        "timeout_terminal_status": (
-                            RowStatus.FAIL if degraded_timeout_finalize else RowStatus.ERROR
-                        ).value,
-                        "timeout_scope": "row_absolute_cap",
-                        "liveness_lease_renewals": liveness_lease_renewals,
-                        "last_progress_signal": last_progress_signal,
-                        "last_progress_signal_type": last_progress_signal_type,
-                        "stall_duration_seconds": round(stall_duration_seconds, 3),
-                        "stall_reason": "absolute_timeout_reached",
-                    }
-                    return self._build_timeout_result_with_trace(
-                        row_id=row_id,
-                        check_id=check.id,
-                        message="Row absolute timeout exceeded",
-                        progress=progress,
-                        timeout_trace=timeout_trace,
-                        matcher_retrieval_trace=matcher_retrieval_trace,
-                    )
-
-                done, _ = await asyncio.wait({task}, timeout=min(poll_interval_seconds, remaining_absolute_seconds))
-                if task in done:
-                    return await task
-
-                progress = self._normalize_timeout_progress_snapshot(self._get_row_runtime_progress(row_id))
-                signal_type = self._detect_lease_progress_signal(previous_progress, progress)
-                extension_decision = self._decide_row_timeout_extension(
-                    timeout_retry_attempt=liveness_lease_renewals,
-                    current_timeout_ms=int(self.config.row_timeout_ms),
-                    progress=progress,
-                    previous_progress_summary=previous_progress_summary,
-                )
-                progress_summary = extension_decision.get("progress_summary")
-                if not isinstance(progress_summary, dict):
-                    progress_summary = self._extract_progress_summary(progress)
-                if isinstance(progress_summary, dict):
-                    previous_progress_summary = dict(progress_summary)
-                previous_progress = progress
-
-                if (
-                    signal_type is not None
-                    or bool(extension_decision.get("allow"))
-                    or self._has_strong_repeat_progress(progress)
-                ):
-                    now = time.monotonic()
-                    last_progress_at = now
-                    last_progress_signal = round(now - started_at, 3)
-                    last_progress_signal_type = signal_type or str(
-                        extension_decision.get("reason") or "progress_heuristic"
-                    )
-                    liveness_lease_renewals += 1
-                    logger.warning(
-                        "row_timeout_lease_renewed",
-                        row_id=row_id,
-                        check_id=check.id,
-                        thread_id=thread_id,
-                        run_id=thread_id,
-                        audit_id=thread_id,
-                        liveness_lease_renewals=liveness_lease_renewals,
-                        last_progress_signal=last_progress_signal,
-                        last_progress_signal_type=last_progress_signal_type,
-                        progress=progress,
-                    )
-                    continue
-
-                now = time.monotonic()
-                elapsed_since_progress_seconds = max(0.0, now - last_progress_at)
-                if elapsed_since_progress_seconds < min(stall_detection_seconds, lease_window_seconds):
-                    continue
-
-                degraded_timeout_finalize = self._should_degrade_timeout_finalize(progress)
-                stall_reason = self._derive_lease_stall_reason(
-                    elapsed_since_progress_seconds=elapsed_since_progress_seconds,
-                    lease_window_seconds=lease_window_seconds,
-                    stall_detection_seconds=stall_detection_seconds,
-                    progress=progress,
-                )
-                timeout_trace = {
-                    "timeout_retry_attempts": 0,
-                    "timeout_progress_snapshot": progress,
-                    "timeout_extended": liveness_lease_renewals > 0,
-                    "timeout_extension_denial_reason": str(extension_decision.get("reason") or stall_reason),
-                    "timeout_degraded_finalize": degraded_timeout_finalize,
-                    "timeout_kind": self._derive_timeout_kind(
-                        progress=progress,
-                        degraded_timeout_finalize=degraded_timeout_finalize,
-                    ),
-                    "timeout_terminal_status": (RowStatus.FAIL if degraded_timeout_finalize else RowStatus.ERROR).value,
-                    "timeout_scope": "row_progress_lease",
-                    "liveness_lease_renewals": liveness_lease_renewals,
-                    "last_progress_signal": last_progress_signal,
-                    "last_progress_signal_type": last_progress_signal_type,
-                    "stall_duration_seconds": round(elapsed_since_progress_seconds, 3),
-                    "stall_reason": stall_reason,
-                }
-                extension_progress_delta = extension_decision.get("progress_delta")
-                if isinstance(extension_progress_delta, dict):
-                    timeout_trace["timeout_extension_progress_delta"] = dict(extension_progress_delta)
-                return self._build_timeout_result_with_trace(
-                    row_id=row_id,
-                    check_id=check.id,
-                    message="Row progress lease expired",
-                    progress=progress,
-                    timeout_trace=timeout_trace,
-                    matcher_retrieval_trace=matcher_retrieval_trace,
-                )
-        finally:
-            if not task.done():
-                task.cancel()
-                with contextlib.suppress(asyncio.CancelledError):
-                    await task
-
-    async def _process_batch_parallel(
-        self,
-        batch_idx: int,
-        start_idx: int,
-        end_idx: int,
-        project_profile: dict[str, Any] | None,
-        checkpoint: BatchCheckpoint,
-        thread_id: str,
-        row_progress_callback: RowProgressCallback | None = None,
-        batch_stats: _ResilienceRetryStats | None = None,
-    ) -> BatchResult:
-        """Process rows in parallel with semaphore-bounded concurrency (FR-188).
-
-        Checkpoint writes and diagnostics dict access are serialized under a lock
-        to preserve correctness under concurrency.
-        """
-        import time
-
-        start_time = time.monotonic()
-        effective_concurrency = self._effective_row_concurrency()
-        logger.info(
-            "row_concurrency_active",
-            effective_concurrency=effective_concurrency,
-            batch_size=end_idx - start_idx,
-            adaptive_mode=bool(self.config.adaptive_row_concurrency),
-            profile=str((project_profile or {}).get("profile") or self.config.run_context.get("active_profile") or ""),
-        )
-        semaphore = asyncio.Semaphore(effective_concurrency)
-        checkpoint_lock = asyncio.Lock()
-        progress_lock = asyncio.Lock()
-        consumer = PostProcessingConsumer()
-        consume_task = asyncio.create_task(consumer.consume())
-        interim_results: list[RowEvaluationResult] = []
-        interim_success_count = 0
-        interim_error_count = 0
-        interim_skipped_count = 0
-
-        async def _guarded_row(idx: int) -> tuple[int, RowEvaluationResult | None, str]:
-            """Wrapper that never lets BudgetExceededError escape gather."""
-            nonlocal interim_success_count, interim_error_count, interim_skipped_count
-            try:
-                _section_id, check = self._checks[idx]
-                row_id = f"{check.id}:row_{idx}"
-                if row_progress_callback is not None:
-                    async with progress_lock:
-                        inflight_result = self._build_inflight_row_progress_result(check, row_id)
-                        maybe_awaitable = row_progress_callback(
-                            batch_idx,
-                            start_idx,
-                            end_idx,
-                            sorted(interim_results, key=lambda item: item.row_id),
-                            interim_success_count,
-                            interim_error_count,
-                            interim_skipped_count,
-                            inflight_result,
-                            "started",
-                        )
-                        if inspect.isawaitable(maybe_awaitable):
-                            await maybe_awaitable
-                result, outcome = await self._process_single_row(
-                    idx=idx,
-                    batch_idx=batch_idx,
-                    project_profile=project_profile,
-                    checkpoint=checkpoint,
-                    thread_id=thread_id,
-                    semaphore=semaphore,
-                    checkpoint_lock=checkpoint_lock,
-                    batch_stats=batch_stats,
-                )
-                if result is not None:
-                    await consumer.put(result)
-                if row_progress_callback is not None:
-                    async with progress_lock:
-                        if result is not None:
-                            interim_results[:] = [res for res in interim_results if res.row_id != result.row_id]
-                            interim_results.append(result)
-                        if outcome == "error":
-                            interim_error_count += 1
-                        elif outcome == "skipped":
-                            interim_skipped_count += 1
-                        else:
-                            interim_success_count += 1
-                        if self._should_emit_row_progress_callback(result, outcome):
-                            maybe_awaitable = row_progress_callback(
-                                batch_idx,
-                                start_idx,
-                                end_idx,
-                                sorted(interim_results, key=lambda item: item.row_id),
-                                interim_success_count,
-                                interim_error_count,
-                                interim_skipped_count,
-                                result,
-                                outcome,
-                            )
-                            if inspect.isawaitable(maybe_awaitable):
-                                await maybe_awaitable
-                return idx, result, outcome
-            except BudgetExceededError as exc:
-                # BudgetExceededError escapes _process_single_row -> _guarded_row
-                # Convert to error result to allow batch to continue
-                logger.warning("parallel_row_budget_exceeded", row_index=idx, error=str(exc))
-                _section_id, check = self._checks[idx]
-                row_id = f"{check.id}:row_{idx}"
-                err_result = self._create_timeout_result(row_id, check.id, f"Budget exceeded: {exc}")
-                await consumer.put(err_result)
-                return idx, err_result, "error"
-            except Exception as exc:
-                logger.warning("parallel_row_error", row_index=idx, error=str(exc))
-                _section_id, check = self._checks[idx]
-                row_id = f"{check.id}:row_{idx}"
-                err_result = self._create_timeout_result(row_id, check.id, f"Parallel row error: {exc}")
-                await consumer.put(err_result)
-                return idx, err_result, "error"
-
-        timed_out = False
-        # P161-E2 TSK-161.11: Non-targeted row skip for single-row shards.
-        # When target_selection specifies a single row index, only create
-        # _guarded_row() coroutines for that index.  Eliminates 58 N/A
-        # iterations per child (60-120 min cumulative waste eliminated).
-        _row_indices = range(start_idx, end_idx)
-        if (
-            self.config.target_selection is not None
-            and len(self.config.target_selection.row_indices) == 1
-            and os.environ.get("VDS_AUDIT_CANONICAL_CHECKLIST_JSON")
-        ):
-            _target_idx = next(iter(self.config.target_selection.row_indices))
-            if start_idx <= _target_idx < end_idx:
-                _row_indices = [_target_idx]
-                logger.info(
-                    "single_row_shard_skip_non_targeted",
-                    targeted_index=_target_idx,
-                    original_range=f"{start_idx}-{end_idx}",
-                    skipped_count=(end_idx - start_idx) - 1,
-                )
-        try:
-            effective_batch_timeout_ms = self.effective_batch_timeout_ms()
-            async with asyncio.timeout(effective_batch_timeout_ms / 1000):
-                gather_results = await asyncio.gather(
-                    *[_guarded_row(idx) for idx in _row_indices],
-                    return_exceptions=False,
-                )
-        except TimeoutError:
-            await consumer.finalize()
-            await consume_task
-            logger.warning(
-                "batch_timeout",
-                batch=batch_idx,
-                completed=0,
-                expected=end_idx - start_idx,
-            )
-            # Create timeout results for all rows in the batch
-            results: list[RowEvaluationResult] = []
-            error_count = 0
-            skipped_count = 0
-            for idx in range(start_idx, end_idx):
-                _section_id, check = self._checks[idx]
-                row_id = f"{check.id}:row_{idx}"
-                if self._should_evaluate_row(idx, check.id, row_id, checkpoint):
-                    progress = self._normalize_timeout_progress_snapshot(self._get_row_runtime_progress(row_id))
-                    degraded = self._should_degrade_timeout_finalize(progress)
-                    timeout_result = self._create_timeout_result(
-                        row_id,
-                        check.id,
-                        "Batch timeout exceeded",
-                        degraded_timeout_finalize=degraded,
-                        progress=progress,
-                    )
-                    timeout_trace = {
-                        "timeout_retry_attempts": 0,
-                        "timeout_progress_snapshot": progress,
-                        "timeout_kind": self._derive_timeout_kind(
-                            progress=progress,
-                            degraded_timeout_finalize=degraded,
-                            batch_timeout_exceeded=True,
-                        ),
-                        "timeout_extended": False,
-                        "timeout_degraded_finalize": degraded,
-                        "timeout_terminal_status": timeout_result.status.value,
-                        "batch_timeout_exceeded": True,
-                    }
-                    self._attach_timeout_progress_telemetry(
-                        timeout_trace=timeout_trace,
-                        progress=progress,
-                        row_id=row_id,
-                        check_id=check.id,
-                    )
-                    timeout_result.retrieval_trace = _sanitize_retrieval_trace(
-                        self._sanitize_timeout_trace_for_app_config_only(
-                            timeout_trace,
-                            progress=progress,
-                            matcher_retrieval_trace=None,
-                        )
-                    )
-                    timeout_result = RowEvaluator._sanitize_app_config_only_result(timeout_result)
-                    timeout_result = RowEvaluator._salvage_app_config_only_timeout_result(timeout_result)
-                    results.append(timeout_result)
-                    error_count += 1
-                elif row_id in checkpoint.results_by_row_id:
-                    try:
-                        preserved = BatchCheckpoint.deserialize_row_result(checkpoint.results_by_row_id[row_id])
-                        results.append(preserved)
-                    except Exception:
-                        pass
-                    skipped_count += 1
-
-            duration_ms = int((time.monotonic() - start_time) * 1000)
-            return BatchResult(
-                batch_index=batch_idx,
-                start_row=start_idx,
-                end_row=end_idx,
-                results=results,
-                duration_ms=duration_ms,
-                success_count=0,
-                error_count=error_count,
-                timed_out=True,
-                skipped_count=skipped_count,
-            )
-
-        # Finalize consumer and collect overlap diagnostics (FR-193)
-        await consumer.finalize()
-        postproc_diagnostics = await consume_task
-
-        # Assemble results in index order
-        results = []
-        success_count = 0
-        error_count = 0
-        skipped_count = 0
-        for _, result, outcome in sorted(gather_results, key=lambda x: x[0]):
-            if result is not None:
-                results.append(result)
-            if outcome == "error":
-                error_count += 1
-            elif outcome == "skipped":
-                skipped_count += 1
-            else:
-                success_count += 1
-
-        # FR-145.2: Post-pass ERROR retry sweep (parallel path mirror of serial path).
-        # Only runs when batch_error_retry_limit > 0 and there are ERROR rows.
-        if not timed_out and error_count > 0 and self.config.batch_error_retry_limit > 0:
-            health_memory: ProviderHealthMemory | None = getattr(self, "_batch_health_memory", None)
-            failover_profiles: list[str] = list(getattr(self, "_batch_failover_profiles", None) or [])
-            if health_memory is not None and failover_profiles:
-                budget = ResilienceBudget(
-                    total_rows=len(results),
-                    error_count=error_count,
-                )
-                retried = await self._retry_error_rows(
-                    results=results,
-                    health_memory=health_memory,
-                    budget=budget,
-                    failover_profiles=failover_profiles,
-                    project_profile=project_profile,
-                    stats=batch_stats,
-                )
-                self._resilience_stats_accumulator.append(batch_stats)
-                new_error_count = sum(1 for r in retried if r.status == RowStatus.ERROR)
-                new_success_count = sum(
-                    1
-                    for r in retried
-                    if r.status != RowStatus.ERROR
-                    and r.row_id in {orig.row_id for orig in results if orig.status == RowStatus.ERROR}
-                )
-                if budget.calls_used > 0:
-                    logger.info(
-                        "resilience_retry_sweep_complete",
-                        batch_index=batch_idx,
-                        errors_before=error_count,
-                        errors_after=new_error_count,
-                        retries_attempted=budget.calls_used,
-                        budget_total=budget.total,
-                    )
-                results = retried
-                error_count = new_error_count
-                success_count += new_success_count
-
-        # FR-13.3b: Check for budget errors in parallel results and raise BatchBudgetExceededError
-        budget_error_results = [
-            r for r in results
-            if "Budget exceeded" in (getattr(r, "reason", "") or "")
-            or "Budget exceeded" in (getattr(r, "error_message", "") or "")
-        ]
-        if budget_error_results:
-            # Get the budget info from first error result
-            budget_result = budget_error_results[0]
-            # Extract row details from the error result
-            row_id = budget_result.row_id
-            check_id = budget_result.check_id
-            # Parse row index from row_id (format: "check_id:row_N")
-            row_index = None
-            if ":row_" in row_id:
-                try:
-                    row_index = int(row_id.split(":row_")[-1]) + 1  # 1-based index
-                except ValueError:
-                    row_index = None
-            context = {
-                "kind": "strict_budget_exceeded",
-                "audit_error_key": "LLM_BUDGET_EXCEEDED",
-                "audit_error_code": AUDIT_ERROR_CODES["LLM_BUDGET_EXCEEDED"].code,
-                "batch_index": batch_idx + 1,
-                "row_index": row_index,
-                "check_id": check_id,
-                "row_id": row_id,
-                "status": "exceeded",
-                "scope": "audit",
-            }
-            raise BatchBudgetExceededError(
-                message=str(budget_result.reason or budget_result.error_message),
-                context=context,
-                partial_results=list(results),
-                success_count=success_count,
-                error_count=error_count,
-                skipped_count=skipped_count,
-            )
-
-        postproc_overlap_ms = int(postproc_diagnostics.get("overlap_seconds", 0.0) * 1000)
-        duration_ms = int((time.monotonic() - start_time) * 1000)
-        return BatchResult(
-            batch_index=batch_idx,
-            start_row=start_idx,
-            end_row=end_idx,
-            results=results,
-            duration_ms=duration_ms,
-            success_count=success_count,
-            error_count=error_count,
-            timed_out=False,
-            skipped_count=skipped_count,
-            postproc_overlap_ms=postproc_overlap_ms,
-        )
-
-    def _get_row_runtime_progress(self, row_id: str) -> dict[str, Any] | None:
-        getter = getattr(self.evaluator, "get_row_runtime_progress", None)
-        if not callable(getter):
-            return None
-        progress = getter(row_id)
-        if not isinstance(progress, dict):
-            return None
-        return dict(progress)
-
-    @staticmethod
-    def _attach_timeout_progress_telemetry(
-        *,
-        timeout_trace: dict[str, Any],
-        progress: dict[str, Any] | None,
-        row_id: str | None = None,
-        check_id: str | None = None,
-    ) -> None:
-        if not isinstance(timeout_trace, dict) or not isinstance(progress, dict):
-            return
-
-        prompt_summary = progress.get("prompt_tool_telemetry_summary")
-        if isinstance(prompt_summary, dict):
-            timeout_trace["prompt_tool_telemetry_summary"] = dict(prompt_summary)
-
-        prompt_payload = progress.get("prompt_tool_telemetry")
-        if isinstance(prompt_payload, dict):
-            timeout_trace["prompt_tool_telemetry"] = dict(prompt_payload)
-
-        interpretation = progress.get("requirement_interpretation")
-        if isinstance(interpretation, dict):
-            timeout_trace["requirement_interpretation"] = dict(interpretation)
-
-        for key in (
-            "route_id",
-            "route_mode",
-            "route_reason",
-            "selection_blocked_reason",
-        ):
-            value = progress.get(key)
-            if value not in (None, False, ""):
-                timeout_trace.setdefault(key, value)
-        for key in (
-            "route_transition_reasons",
-            "project_artifact_provenance_summary",
-            "cross_repo_evidence_refs",
-            "config_repo_companion_refs",
-        ):
-            value = progress.get(key)
-            if isinstance(value, list) and value:
-                timeout_trace.setdefault(key, list(value))
-        for key in (
-            "docs_intent_utilization",
-            "code_ranking_diagnostics",
-            "project_artifact_readiness",
-        ):
-            value = progress.get(key)
-            if isinstance(value, dict) and value:
-                timeout_trace.setdefault(key, dict(value))
-
-        timeout_kind = str(timeout_trace.get("timeout_kind") or "").strip()
-        timeout_terminal_status = str(timeout_trace.get("timeout_terminal_status") or "").strip().upper() or "ERROR"
-        raw_refs = progress.get("evidence_refs")
-        fallback_anchor_count = len(raw_refs) if isinstance(raw_refs, list) else 0
-        if "grounding_validation" not in timeout_trace:
-            timeout_trace["grounding_validation"] = {
-                "complete": False,
-                "timeout_incomplete": True,
-                "timeout_kind": timeout_kind or False,
-                "total_anchors": fallback_anchor_count,
-                "valid_count": 0,
-                "invalid_count": 0,
-                "all_grounded": False,
-                "grounding_ratio": 0.0,
-                "strict_mode": False,
-                "strict_mode_zero_valid_error": False,
-                "invalid_ref_values": [],
-                "gap_signal": False,
-            }
-        if "evidence_gap_diagnostics" not in timeout_trace:
-            interpretation_mapping = interpretation if isinstance(interpretation, dict) else {}
-            readiness_mapping = (
-                dict(progress.get("project_artifact_readiness"))
-                if isinstance(progress.get("project_artifact_readiness"), dict)
-                else {}
-            )
-            diagnostics: dict[str, Any] = {
-                "applied": False,
-                "classification": False,
-                "gap_type": False,
-                "status_considered": timeout_terminal_status,
-                "dominant_unverified_reason": False,
-                "verified_true_count": 0,
-                "verified_false_count": fallback_anchor_count,
-                "incomplete_due_to_timeout": True,
-                "timeout_kind": timeout_kind or False,
-            }
-            if bool(interpretation_mapping.get("project_scope_required")):
-                authoritative_ready = bool(readiness_mapping.get("authoritative_artifact_ready"))
-                missing_artifacts = list(readiness_mapping.get("missing_artifact_types") or [])
-                if readiness_mapping and (not authoritative_ready or bool(missing_artifacts)):
-                    diagnostics.update(
-                        {
-                            "applied": True,
-                            "classification": "confirmed_gap",
-                            "gap_type": "authoritative_project_scope_artifact_missing",
-                        }
-                    )
-            timeout_trace["evidence_gap_diagnostics"] = diagnostics
-
-        skill_policy_preview = progress.get("skill_policy_preview")
-        if not isinstance(skill_policy_preview, dict):
-            return
-
-        def _safe_int(value: Any) -> int:
-            try:
-                return int(value or 0)
-            except (TypeError, ValueError):
-                return 0
-
-        skill_policy_timeout = dict(skill_policy_preview)
-        skill_policy_timeout.setdefault("incomplete_due_to_timeout", True)
-        if isinstance(prompt_summary, dict):
-            skill_policy_timeout["observed_skill_calls"] = max(
-                _safe_int(skill_policy_timeout.get("observed_skill_calls")),
-                _safe_int(prompt_summary.get("skill_calls")),
-            )
-            skill_policy_timeout["observed_skill_execution_calls"] = max(
-                _safe_int(skill_policy_timeout.get("observed_skill_execution_calls")),
-                _safe_int(prompt_summary.get("skill_execution_calls")),
-            )
-            skill_policy_timeout["observed_skill_effective_calls"] = max(
-                _safe_int(skill_policy_timeout.get("observed_skill_effective_calls")),
-                _safe_int(prompt_summary.get("skill_effective_calls")),
-            )
-        timeout_trace["skill_policy_retry"] = skill_policy_timeout
-        logger.info(
-            "row_skill_policy_timeout_snapshot",
-            row_id=row_id,
-            check_id=check_id,
-            skill_metrics_scope="timeout_progress_snapshot",
-            required=bool(skill_policy_timeout.get("required")),
-            policy_mode=str(skill_policy_timeout.get("policy_mode") or "unknown"),
-            strict_require_effective_skill=bool(skill_policy_timeout.get("strict_require_effective_skill")),
-            enforcement_enabled=bool(skill_policy_timeout.get("enforcement_enabled")),
-            incomplete_due_to_timeout=True,
-            observed_skill_calls=_safe_int(skill_policy_timeout.get("observed_skill_calls")),
-            observed_skill_execution_calls=_safe_int(skill_policy_timeout.get("observed_skill_execution_calls")),
-            observed_skill_effective_calls=_safe_int(skill_policy_timeout.get("observed_skill_effective_calls")),
-        )
-
-    def effective_row_timeout_ceiling_ms(self) -> int:
-        """Return maximum per-attempt row timeout reachable under extension policy."""
-        if self._lease_timeout_mode_enabled():
-            return max(1, int(self.config.row_absolute_timeout_ms))
-        base_timeout_ms = max(1, int(self.config.row_timeout_ms))
-        if not bool(self.config.row_timeout_progress_extension_enabled):
-            return base_timeout_ms
-
-        extension_steps = self._effective_timeout_extension_steps(base_timeout_ms)
-        if extension_steps <= 0:
-            return base_timeout_ms
-
-        timeout_ms = base_timeout_ms
-        for attempt in range(extension_steps):
-            next_timeout_ms = self._next_row_timeout_ms(timeout_ms, attempt)
-            if next_timeout_ms <= timeout_ms:
-                break
-            timeout_ms = next_timeout_ms
-        return timeout_ms
-
-    def effective_row_timeout_envelope_ms(self) -> int:
-        """Return cumulative per-row timeout envelope across all retry attempts.
-
-        Row timeout retries currently re-run row evaluation with progressively larger
-        per-attempt timeouts. The total elapsed time can therefore exceed the final
-        per-attempt timeout ceiling. Batch timeout budgeting must account for this
-        cumulative envelope to avoid preempting rows that are still making progress.
-        """
-        if self._lease_timeout_mode_enabled():
-            return max(1, int(self.config.row_absolute_timeout_ms))
-        current_timeout_ms = max(1, int(self.config.row_timeout_ms))
-        envelope_ms = current_timeout_ms
-
-        if not bool(self.config.row_timeout_progress_extension_enabled):
-            return envelope_ms
-
-        extension_steps = self._effective_timeout_extension_steps(current_timeout_ms)
-        if extension_steps <= 0:
-            return envelope_ms
-
-        for attempt in range(extension_steps):
-            next_timeout_ms = self._next_row_timeout_ms(current_timeout_ms, attempt)
-            if next_timeout_ms <= current_timeout_ms:
-                break
-            envelope_ms += next_timeout_ms
-            current_timeout_ms = next_timeout_ms
-
-        return envelope_ms
-
-    def effective_batch_timeout_ms(self) -> int:
-        """Return effective batch timeout with cumulative progress-aware retry headroom."""
-        if self.config.batch_timeout_ms is not None:
-            return int(self.config.batch_timeout_ms)
-        return int(self.config.batch_size * self.effective_row_timeout_envelope_ms() + self.config.batch_overhead_ms)
-
-    def _lease_timeout_mode_enabled(self) -> bool:
-        return int(self.config.row_progress_lease_seconds or 0) > 0
-
-    def _clear_row_runtime_progress(self, row_id: str) -> None:
-        clearer = getattr(self.evaluator, "clear_row_runtime_progress", None)
-        if callable(clearer):
-            clearer(row_id)
-
-    def _effective_timeout_extension_steps(self, base_timeout_ms: int) -> int:
-        """Return extension steps to budget when progress-based cap override is enabled."""
-        retry_attempts = max(0, int(self.config.row_timeout_progress_retry_attempts))
-        if retry_attempts <= 0:
-            return 0
-
-        max_timeout_ms = max(int(self.config.row_timeout_ms), int(self.config.row_timeout_progress_max_ms))
-        if max_timeout_ms <= base_timeout_ms:
-            return retry_attempts
-
-        extension_ms = max(1, int(self.config.row_timeout_progress_extension_ms))
-        steps_to_max = math.ceil((max_timeout_ms - base_timeout_ms) / extension_ms)
-
-        # Runtime can continue beyond retry cap when telemetry still shows progress;
-        # batch budgeting should include that headroom to avoid premature batch timeout.
-        return max(retry_attempts, steps_to_max)
-
-    def _next_row_timeout_ms(self, current_timeout_ms: int, timeout_retry_attempt: int) -> int:
-        max_timeout_ms = max(int(self.config.row_timeout_ms), int(self.config.row_timeout_progress_max_ms))
-        base_extension_ms = max(1, int(self.config.row_timeout_progress_extension_ms))
-
-        # Keep increment size stable across retries so progressing rows can
-        # accumulate enough wall-clock budget before finalization.
-        extension_ms = base_extension_ms
-
-        return min(max_timeout_ms, current_timeout_ms + extension_ms)
-
-    @staticmethod
-    def _coerce_progress_counter(value: Any) -> int:
-        if isinstance(value, bool):
-            return int(value)
-        if isinstance(value, int | float):
-            return max(0, int(value))
-        if isinstance(value, str):
-            try:
-                return max(0, int(float(value.strip())))
-            except ValueError:
-                return 0
-        return 0
-
-    @staticmethod
-    def _coerce_progress_timestamp(value: Any) -> float:
-        if isinstance(value, bool):
-            return 0.0
-        if isinstance(value, int | float):
-            return float(value)
-        if isinstance(value, str):
-            try:
-                return float(value.strip())
-            except ValueError:
-                return 0.0
-        return 0.0
-
-    def _derive_progress_summary_from_snapshot(
-        self,
-        progress: dict[str, Any] | None,
-        *,
-        include_step_estimate: bool = True,
-    ) -> dict[str, int] | None:
-        if not isinstance(progress, dict):
-            return None
-
-        prompt_summary = progress.get("prompt_tool_telemetry_summary")
-        skill_preview = progress.get("skill_policy_preview")
-        summary: dict[str, int] = dict.fromkeys(_TIMEOUT_PROGRESS_COUNTER_KEYS, 0)
-        signal_detected = False
-
-        if isinstance(prompt_summary, dict):
-            for key in _TIMEOUT_PROGRESS_COUNTER_KEYS:
-                value = self._coerce_progress_counter(prompt_summary.get(key))
-                summary[key] = max(summary[key], value)
-                signal_detected = signal_detected or value > 0
-
-        if isinstance(skill_preview, dict):
-            observed_skill_calls = self._coerce_progress_counter(skill_preview.get("observed_skill_calls"))
-            observed_skill_execution_calls = self._coerce_progress_counter(
-                skill_preview.get("observed_skill_execution_calls")
-            )
-            observed_skill_effective_calls = self._coerce_progress_counter(
-                skill_preview.get("observed_skill_effective_calls")
-            )
-            summary["skill_calls"] = max(summary["skill_calls"], observed_skill_calls)
-            summary["skill_execution_calls"] = max(summary["skill_execution_calls"], observed_skill_execution_calls)
-            summary["skill_effective_calls"] = max(summary["skill_effective_calls"], observed_skill_effective_calls)
-            signal_detected = signal_detected or any(
-                value > 0
-                for value in (
-                    observed_skill_calls,
-                    observed_skill_execution_calls,
-                    observed_skill_effective_calls,
-                )
-            )
-
-        if include_step_estimate:
-            tool_calls_from_steps = self._coerce_progress_counter(progress.get("steps_executed"))
-            summary["tool_calls"] = max(summary["tool_calls"], tool_calls_from_steps)
-            signal_detected = signal_detected or tool_calls_from_steps > 0
-
-        if signal_detected or isinstance(prompt_summary, dict) or isinstance(skill_preview, dict):
-            return summary
-        return None
-
-    def _normalize_timeout_progress_snapshot(self, progress: dict[str, Any] | None) -> dict[str, Any] | None:
-        if not isinstance(progress, dict):
-            return progress
-
-        normalized = dict(progress)
-        original_prompt_summary = progress.get("prompt_tool_telemetry_summary")
-        original_prompt_payload = progress.get("prompt_tool_telemetry")
-        original_skill_preview = progress.get("skill_policy_preview")
-        summary = self._derive_progress_summary_from_snapshot(normalized, include_step_estimate=True)
-        if not isinstance(summary, dict):
-            return normalized
-
-        normalized["prompt_tool_telemetry_summary"] = dict(summary)
-        normalized["timeout_progress_summary_source"] = (
-            "explicit" if isinstance(original_prompt_summary, dict) else "derived"
-        )
-        normalized["timeout_progress_has_explicit_prompt_telemetry"] = isinstance(original_prompt_payload, dict)
-        normalized["timeout_progress_has_explicit_skill_preview"] = isinstance(original_skill_preview, dict)
-
-        prompt_payload = normalized.get("prompt_tool_telemetry")
-        if isinstance(prompt_payload, dict):
-            merged_payload = dict(prompt_payload)
-            merged_payload["event_tool_calls_completed"] = max(
-                self._coerce_progress_counter(merged_payload.get("event_tool_calls_completed")),
-                self._coerce_progress_counter(summary.get("tool_calls")),
-            )
-            merged_payload["event_skill_tool_calls"] = max(
-                self._coerce_progress_counter(merged_payload.get("event_skill_tool_calls")),
-                self._coerce_progress_counter(summary.get("skill_calls")),
-            )
-            merged_payload["event_skill_execution_tool_calls"] = max(
-                self._coerce_progress_counter(merged_payload.get("event_skill_execution_tool_calls")),
-                self._coerce_progress_counter(summary.get("skill_execution_calls")),
-            )
-            merged_payload["event_skill_effective_tool_calls"] = max(
-                self._coerce_progress_counter(merged_payload.get("event_skill_effective_tool_calls")),
-                self._coerce_progress_counter(summary.get("skill_effective_calls")),
-            )
-            merged_payload["event_skill_execution_effective_tool_calls"] = max(
-                self._coerce_progress_counter(merged_payload.get("event_skill_execution_effective_tool_calls")),
-                self._coerce_progress_counter(summary.get("skill_effective_calls")),
-            )
-            normalized["prompt_tool_telemetry"] = merged_payload
-
-        skill_preview = normalized.get("skill_policy_preview")
-        if isinstance(skill_preview, dict):
-            merged_skill_preview = dict(skill_preview)
-            merged_skill_preview["observed_skill_calls"] = max(
-                self._coerce_progress_counter(merged_skill_preview.get("observed_skill_calls")),
-                self._coerce_progress_counter(summary.get("skill_calls")),
-            )
-            merged_skill_preview["observed_skill_execution_calls"] = max(
-                self._coerce_progress_counter(merged_skill_preview.get("observed_skill_execution_calls")),
-                self._coerce_progress_counter(summary.get("skill_execution_calls")),
-            )
-            merged_skill_preview["observed_skill_effective_calls"] = max(
-                self._coerce_progress_counter(merged_skill_preview.get("observed_skill_effective_calls")),
-                self._coerce_progress_counter(summary.get("skill_effective_calls")),
-            )
-            normalized["skill_policy_preview"] = merged_skill_preview
-
-        return normalized
-
-    def _extract_progress_summary(self, progress: dict[str, Any] | None) -> dict[str, int] | None:
-        summary = self._derive_progress_summary_from_snapshot(progress, include_step_estimate=False)
-        if isinstance(summary, dict):
-            return dict(summary)
-        return None
-
-    def _has_strong_repeat_progress(self, progress: dict[str, Any] | None) -> bool:
-        """Allow bounded repeat extension when repeated snapshots still show strong work."""
-        if not isinstance(progress, dict):
-            return False
-        if not bool(progress.get("effective_progress")) or not self._progress_has_grounding_signal(progress):
-            return False
-        # Prevent stale pre-LLM "completed" snapshots from repeatedly extending timeout.
-        # Strong repeat extension is only valid while tool-first loop itself is still timing out.
-        termination_reason = str(progress.get("termination_reason") or "").strip().lower()
-        if termination_reason and termination_reason != "timeout":
-            return False
-        steps_executed = self._coerce_progress_counter(progress.get("steps_executed"))
-        steps_unique_tools = self._coerce_progress_counter(progress.get("steps_unique_tools"))
-        repetition_raw = progress.get("step_repetition_rate")
-        repetition_rate = float(repetition_raw) if repetition_raw is not None else 1.0
-        return steps_executed >= 3 and steps_unique_tools >= 2 and repetition_rate <= 0.6
-
-    @staticmethod
-    def _compute_progress_delta(
-        previous_summary: dict[str, int] | None,
-        current_summary: dict[str, int] | None,
-    ) -> dict[str, int] | None:
-        if not isinstance(previous_summary, dict) or not isinstance(current_summary, dict):
-            return None
-        delta: dict[str, int] = {}
-        for key in _TIMEOUT_PROGRESS_COUNTER_KEYS:
-            previous_value = max(0, int(previous_summary.get(key, 0)))
-            current_value = max(0, int(current_summary.get(key, 0)))
-            delta[key] = max(0, current_value - previous_value)
-        return delta
-
-    def _decide_row_timeout_extension(
-        self,
-        *,
-        timeout_retry_attempt: int,
-        current_timeout_ms: int,
-        progress: dict[str, Any] | None,
-        previous_progress_summary: dict[str, int] | None,
-    ) -> dict[str, Any]:
-        decision: dict[str, Any] = {
-            "allow": False,
-            "reason": "not_allowed",
-            "progress_summary": None,
-            "progress_delta": None,
-        }
-        if not bool(self.config.row_timeout_progress_extension_enabled):
-            decision["reason"] = "extension_disabled"
-            return decision
-        retry_attempts = max(0, int(self.config.row_timeout_progress_retry_attempts))
-        if retry_attempts <= 0:
-            decision["reason"] = "attempt_cap_reached"
-            return decision
-        extension_cap_reached = timeout_retry_attempt >= retry_attempts
-        if current_timeout_ms >= max(int(self.config.row_timeout_ms), int(self.config.row_timeout_progress_max_ms)):
-            decision["reason"] = "max_timeout_reached"
-            return decision
-        if not isinstance(progress, dict):
-            decision["reason"] = "missing_progress_snapshot"
-            return decision
-        if not bool(progress.get("effective_progress")):
-            decision["reason"] = "ineffective_progress_snapshot"
-            return decision
-        if not self._progress_has_grounding_signal(progress):
-            decision["reason"] = "missing_grounding_signal"
-            return decision
-
-        termination_reason = str(progress.get("termination_reason") or "").strip().lower()
-        if timeout_retry_attempt < 1 and termination_reason and termination_reason != "timeout":
-            decision["reason"] = "termination_reason_not_timeout"
-            return decision
-
-        if timeout_retry_attempt >= 1:
-            # Repeat extensions remain bounded, but should continue when
-            # telemetry still shows meaningful work.
-            repetition_raw = progress.get("step_repetition_rate")
-            repetition_rate = float(repetition_raw) if repetition_raw is not None else 1.0
-            if repetition_rate >= 0.7:
-                decision["reason"] = "high_step_repetition_rate"
-                return decision
-
-            current_summary = self._extract_progress_summary(progress)
-            decision["progress_summary"] = current_summary
-            if current_summary is not None and previous_progress_summary is not None:
-                progress_delta = self._compute_progress_delta(previous_progress_summary, current_summary)
-                decision["progress_delta"] = progress_delta
-                if isinstance(progress_delta, dict) and max(progress_delta.values(), default=0) > 0:
-                    decision["allow"] = True
-                    decision["reason"] = (
-                        "repeat_extension_progress_delta_beyond_cap"
-                        if extension_cap_reached
-                        else "repeat_extension_progress_delta"
-                    )
-                    return decision
-
-            # Strong grounded progress may still justify repeat extensions even
-            # when prompt telemetry counters remain stationary across retries.
-            if self._has_strong_repeat_progress(progress):
-                decision["allow"] = True
-                if current_summary is None:
-                    decision["progress_summary"] = {
-                        "tool_calls": self._coerce_progress_counter(progress.get("steps_executed")),
-                        "skill_calls": 0,
-                        "skill_execution_calls": 0,
-                        "skill_effective_calls": 0,
-                    }
-                    decision["reason"] = (
-                        "repeat_extension_strong_progress_no_prompt_summary_beyond_cap"
-                        if extension_cap_reached
-                        else "repeat_extension_strong_progress_no_prompt_summary"
-                    )
-                elif previous_progress_summary is None:
-                    decision["reason"] = (
-                        "repeat_extension_strong_progress_missing_previous_summary_beyond_cap"
-                        if extension_cap_reached
-                        else "repeat_extension_strong_progress_missing_previous_summary"
-                    )
-                else:
-                    decision["reason"] = (
-                        "repeat_extension_strong_progress_stationary_summary_beyond_cap"
-                        if extension_cap_reached
-                        else "repeat_extension_strong_progress_stationary_summary"
-                    )
-                return decision
-
-            if current_summary is None:
-                decision["reason"] = (
-                    "missing_prompt_telemetry_for_repeat_extension_beyond_cap"
-                    if extension_cap_reached
-                    else "missing_prompt_telemetry_for_repeat_extension"
-                )
-            elif previous_progress_summary is None:
-                decision["reason"] = (
-                    "missing_previous_progress_summary_for_repeat_extension_beyond_cap"
-                    if extension_cap_reached
-                    else "missing_previous_progress_summary_for_repeat_extension"
-                )
-            else:
-                decision["reason"] = (
-                    "no_delta_since_last_timeout_beyond_cap" if extension_cap_reached else "no_delta_since_last_timeout"
-                )
-                decision["progress_delta"] = self._compute_progress_delta(previous_progress_summary, current_summary)
-            return decision
-
-        if extension_cap_reached:
-            decision["reason"] = "attempt_cap_reached"
-            return decision
-        decision["allow"] = True
-        decision["reason"] = "first_extension_allowed"
-        decision["progress_summary"] = self._extract_progress_summary(progress)
-        return decision
-
-    def _progress_has_grounding_signal(self, progress: dict[str, Any] | None) -> bool:
-        if not isinstance(progress, dict):
-            return False
-        evidence_refs_count = progress.get("evidence_refs_count")
-        if isinstance(evidence_refs_count, bool):
-            return bool(evidence_refs_count)
-        if isinstance(evidence_refs_count, int | float):
-            return evidence_refs_count > 0
-        if isinstance(evidence_refs_count, str):
-            try:
-                return float(evidence_refs_count) > 0
-            except ValueError:
-                return False
-        if isinstance(evidence_refs_count, list | tuple | set | dict):
-            return len(evidence_refs_count) > 0
-        return False
-
-    @staticmethod
-    def _progress_has_recorded_or_verified_grounding(progress: dict[str, Any] | None) -> bool:
-        if not isinstance(progress, dict):
-            return False
-        if bool(progress.get("record_evidence_refs_contract_met")):
-            return True
-        raw_payload = progress.get("evidence_ref_verifications")
-        entries: list[dict[str, Any]] = []
-        if isinstance(raw_payload, dict):
-            for ref_value, payload in raw_payload.items():
-                if isinstance(payload, dict):
-                    merged = dict(payload)
-                    merged.setdefault("ref_value", str(ref_value or "").strip())
-                    entries.append(merged)
-        elif isinstance(raw_payload, list):
-            entries = [entry for entry in raw_payload if isinstance(entry, dict)]
-        for entry in entries:
-            if not bool(entry.get("verified")):
-                continue
-            excerpt = str(entry.get("excerpt") or "").strip()
-            if excerpt:
-                return True
-        return False
-
-    def _should_degrade_timeout_finalize(self, progress: dict[str, Any] | None) -> bool:
-        """Return whether timeout should degrade to grounded terminal status (FR-128)."""
-        if not isinstance(progress, dict):
-            return False
-        if not bool(progress.get("effective_progress")):
-            return False
-
-        has_grounding_signal = self._progress_has_grounding_signal(progress)
-        if not has_grounding_signal:
-            return False
-
-        termination_reason = str(progress.get("termination_reason") or "").strip().lower()
-        if termination_reason == "completed":
-            if self._progress_has_recorded_or_verified_grounding(progress):
-                return True
-
-            explicit_prompt_summary = progress.get("prompt_tool_telemetry_summary")
-            has_explicit_prompt_summary = progress.get("timeout_progress_summary_source") == "explicit" and isinstance(
-                explicit_prompt_summary, dict
-            )
-            if has_explicit_prompt_summary and any(
-                self._coerce_progress_counter(explicit_prompt_summary.get(key)) > 0
-                for key in _TIMEOUT_PROGRESS_COUNTER_KEYS
-            ):
-                return True
-
-            explicit_prompt_payload = progress.get("prompt_tool_telemetry")
-            if (
-                bool(progress.get("timeout_progress_has_explicit_prompt_telemetry"))
-                and isinstance(explicit_prompt_payload, dict)
-                and any(
-                    self._coerce_progress_counter(explicit_prompt_payload.get(key)) > 0
-                    for key in (
-                        "event_tool_calls_completed",
-                        "event_skill_tool_calls",
-                        "event_skill_execution_tool_calls",
-                        "event_skill_effective_tool_calls",
-                        "event_skill_execution_effective_tool_calls",
-                    )
-                )
-            ):
-                return True
-
-            explicit_skill_preview = progress.get("skill_policy_preview")
-            if (
-                bool(progress.get("timeout_progress_has_explicit_skill_preview"))
-                and isinstance(explicit_skill_preview, dict)
-                and any(
-                    self._coerce_progress_counter(explicit_skill_preview.get(key)) > 0
-                    for key in (
-                        "observed_skill_calls",
-                        "observed_skill_execution_calls",
-                        "observed_skill_effective_calls",
-                    )
-                )
-            ):
-                return True
-
-            # Phase 130: when the operator explicitly granted additional retry
-            # budget but the tool loop already reports a completed snapshot,
-            # finalize as grounded FAIL instead of a generic ERROR.
-            return max(0, int(self.config.row_timeout_progress_retry_attempts)) > 1
-
-        return True
-
-    @staticmethod
-    def _derive_timeout_kind(
-        *,
-        progress: dict[str, Any] | None,
-        degraded_timeout_finalize: bool,
-        batch_timeout_exceeded: bool = False,
-        absolute_timeout_exceeded: bool = False,
-    ) -> str:
-        if degraded_timeout_finalize:
-            return TimeoutKind.TIMEOUT_WITH_GROUNDING.value
-        if not isinstance(progress, dict):
-            return TimeoutKind.TIMEOUT_NO_PROGRESS.value
-        termination_reason = str(progress.get("termination_reason") or "").strip().lower()
-        if batch_timeout_exceeded:
-            return TimeoutKind.TIMEOUT_ABSOLUTE_CAP.value
-        if absolute_timeout_exceeded:
-            return TimeoutKind.TIMEOUT_ABSOLUTE_CAP.value
-        if termination_reason in {"idle_after_tool", "post_tool_idle", "idle_post_tool"}:
-            return TimeoutKind.TIMEOUT_IDLE_AFTER_TOOL.value
-        if termination_reason in {"tool_churn", "agent_tool_churn", "churn", "stagnation"}:
-            return TimeoutKind.TIMEOUT_CHURN_DETECTED.value
-        if termination_reason in {"provider_stall", "stream_stall"}:
-            return TimeoutKind.TIMEOUT_PROVIDER_STALL.value
-        return TimeoutKind.TIMEOUT_NO_PROGRESS.value
-
-    @staticmethod
-    def _build_timeout_fallback_verification_index(progress: dict[str, Any] | None) -> dict[str, dict[str, Any]]:
-        if not isinstance(progress, dict):
-            return {}
-        raw_payload = progress.get("evidence_ref_verifications")
-        verification_index: dict[str, dict[str, Any]] = {}
-        if isinstance(raw_payload, dict):
-            for key, value in raw_payload.items():
-                ref_value = str(key or "").strip()
-                if not ref_value:
-                    continue
-                verification_index[ref_value] = value if isinstance(value, dict) else {"verified": bool(value)}
-            return verification_index
-        if not isinstance(raw_payload, list):
-            return {}
-        for entry in raw_payload:
-            if not isinstance(entry, dict):
-                continue
-            ref_value = str(entry.get("ref_value") or entry.get("ref") or entry.get("evidence_ref") or "").strip()
-            if not ref_value:
-                continue
-            verification_index[ref_value] = dict(entry)
-        return verification_index
-
-    def _create_timeout_result(
-        self,
-        row_id: str,
-        check_id: str,
-        message: str,
-        *,
-        degraded_timeout_finalize: bool = False,
-        progress: dict[str, Any] | None = None,
-    ) -> RowEvaluationResult:
-        """Create timeout result (ERROR default, degraded FAIL when grounded progress exists)."""
-        degraded = bool(degraded_timeout_finalize)
-        evidence_anchors: list[EvidenceAnchor] = []
-        if degraded and isinstance(progress, dict):
-            verification_index = self._build_timeout_fallback_verification_index(progress)
-            raw_refs = progress.get("evidence_refs")
-            if isinstance(raw_refs, list):
-                for ref in raw_refs[:20]:
-                    ref_value = str(ref or "").strip()
-                    if not ref_value:
-                        continue
-                    verification_payload = verification_index.get(ref_value, {})
-                    excerpt = str(verification_payload.get("excerpt") or "").strip()
-                    verified = bool(verification_payload.get("verified")) and bool(excerpt)
-                    verification_reason = VERIFICATION_REASON_FALLBACK_REF_INHERITED
-                    if verified:
-                        raw_reason = str(verification_payload.get("verification_reason") or "").strip()
-                        verification_reason = normalize_verification_reason(
-                            raw_reason or VERIFICATION_REASON_EXCERPT_VERIFIED_IN_CONTEXT,
-                            verified=True,
-                            excerpt=excerpt,
-                        )
-                    evidence_anchors.append(
-                        EvidenceAnchor(
-                            ref_type=(
-                                "url"
-                                if "://" in ref_value
-                                else "config"
-                                if RowEvaluator._is_config_ref_value(ref_value)
-                                else "doc_path"
-                            ),
-                            ref_value=ref_value,
-                            excerpt=excerpt if verified else "",
-                            verified=verified,
-                            verification_reason=verification_reason,
-                        )
-                    )
-        return RowEvaluationResult(
-            row_id=row_id,
-            check_id=check_id,
-            status=RowStatus.FAIL if degraded else RowStatus.ERROR,
-            score=0.0,
-            score_breakdown=ScoreBreakdown.compute(0.0),
-            reason=(
-                "Đánh giá kết thúc suy giảm sau timeout có tiến triển hiệu quả." if degraded else "Đánh giá bị timeout."
-            ),
-            reasoning="",
-            finding=(
-                "Có tín hiệu grounding tối thiểu trước timeout; trả về kết quả FAIL suy giảm."
-                if degraded
-                else "Không thể hoàn thành đánh giá do timeout."
-            ),
-            evidence_anchors=evidence_anchors,
-            provenance=RowProvenance(
-                row_llm_mode=self.evaluator.config.mode.value,
-                protocol=None,
-                model=None,
-                template_hash=self.evaluator.template_hash,
-                rubric_version="",
-                evidence_hash=self.evaluator.evidence_hash or "",
-                evaluated_at=datetime.now(UTC),
-                cache_hit=False,
-            ),
-            error_message=message,
-        )
-
-    def _get_checkpoint_path(self, thread_id: str) -> Path:
-        """Get checkpoint file path."""
-        if self.config.checkpoint_dir:
-            return self.config.checkpoint_dir / f"batch-checkpoint-{thread_id}.json"
-        return Path(f"/tmp/vds-audit-checkpoint-{thread_id}.json")
-
-    def get_checkpoint_path(self, thread_id: str) -> Path:
-        """Return the checkpoint file path for a given thread ID (FR-278).
-
-        Public accessor for callers that need to bind checkpoint location
-        into workflow-summary.json and run-history.json entries.
-        """
-        return self._get_checkpoint_path(thread_id)
-
-    @staticmethod
-    def _extract_prompt_level_telemetry(result: Any) -> dict[str, int]:
-        """Phase 166: Extract prompt-level retry/failover counts from a row result."""
-        counts: dict[str, int] = {
-            "prompt_retry_attempts": 0,
-            "prompt_retry_recoveries": 0,
-            "prompt_failover_attempts": 0,
-            "prompt_failover_recoveries": 0,
-            "prompt_failover_exhausted": 0,
-            "synthesis_fallback_count": 0,
-        }
-        if result is None:
-            return counts
-        provenance = getattr(result, "provenance", None) or {}
-        if isinstance(provenance, dict):
-            retry_meta = provenance.get("retry_metadata") or {}
-            if isinstance(retry_meta, dict):
-                provider_retry = retry_meta.get("provider_retry") or {}
-                if isinstance(provider_retry, dict):
-                    counts["prompt_retry_attempts"] += max(0, int(provider_retry.get("attempts_executed", 0) or 0))
-                    if provider_retry.get("recovered"):
-                        counts["prompt_retry_recoveries"] += 1
-                failover_tel = retry_meta.get("prompt_failover_telemetry") or {}
-                if isinstance(failover_tel, dict) and failover_tel.get("provider_failover_attempted"):
-                    counts["prompt_failover_attempts"] += 1
-                    timeout_kind = str(failover_tel.get("timeout_kind") or "")
-                    if timeout_kind == "timeout_failover_exhausted":
-                        counts["prompt_failover_exhausted"] += 1
-                    elif failover_tel.get("provider_failover_final_provider"):
-                        counts["prompt_failover_recoveries"] += 1
-                # Check for synthesis fallback
-                fallback_cause = retry_meta.get("fallback_cause") or provenance.get("fallback_cause")
-                if fallback_cause:
-                    counts["synthesis_fallback_count"] += 1
-        return counts
-
-
-class PostProcessingConsumer:
-    """Incremental row result consumer for streaming post-processing (FR-193).
-
-    Consumes ``RowEvaluationResult`` items from an asyncio queue as they arrive,
-    accumulating partial scores and computing overlap timing diagnostics.  A
-    ``None`` sentinel signals the end of the stream.
-    """
-
-    def __init__(self) -> None:
-        self._queue: asyncio.Queue[RowEvaluationResult | None] = asyncio.Queue()
-        self._partial_scores: list[float] = []
-        self._processed_count: int = 0
-        self._start_time: float = 0.0
-        self._overlap_seconds: float = 0.0
-
-    async def put(self, result: RowEvaluationResult) -> None:
-        """Enqueue a completed row result for incremental processing."""
-        await self._queue.put(result)
-
-    async def finalize(self) -> None:
-        """Signal end-of-stream by enqueuing a ``None`` sentinel."""
-        await self._queue.put(None)
-
-    async def consume(self) -> dict[str, Any]:
-        """Consume results until sentinel, return timing diagnostics."""
-        import time
-
-        self._start_time = time.monotonic()
-        while True:
-            item = await self._queue.get()
-            if item is None:
-                break
-            self._processed_count += 1
-            self._partial_scores.append(item.score)
-        end_time = time.monotonic()
-        self._overlap_seconds = end_time - self._start_time if self._start_time else 0.0
-        return {
-            "overlap_seconds": self._overlap_seconds,
-            "processed_count": self._processed_count,
-            "partial_scores": list(self._partial_scores),
-        }